Gear case working condition classification method and system

Abstract

The invention relates to a gear box working condition classification method and system. The method comprises the following steps: encoding and compressing a collected multi-mode sensor signal of a gear box by using a side end server, and uploading the encoded and compressed signal to a cloud server; performing data decoding on the cloud server to obtain a multi-modal sensor signal; deploying a teacher model in the cloud server, and carrying out working condition classification training on the teacher model based on the multi-modal sensor signal; deploying a student model at a side end server, carrying out working condition classification training on the student model based on a multi-modal sensor signal, and guiding the training of the student model by utilizing a teacher model based on a knowledge distillation mechanism; and utilizing the trained student model to predict and classify the working condition of the gearbox. Based on a knowledge distillation mechanism, a teacher model deployed at a cloud end is utilized to guide training of a student model deployed at an edge end, the real-time requirement of a working condition classification task and limitation of insufficient computing resources of the edge end are fully considered, and high-precision, low-delay and low-cost working condition classification is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of gearboxes, and in particular to a method and system for classifying gearbox operating conditions. Background Art

[0002] As a crucial component of mechanical equipment, the operating status of a gearbox is not only directly related to the overall performance of the equipment, but also affects the reliability and service life of the system. The stable operation of gearboxes is crucial in a variety of fields, including engineering machinery, automobile manufacturing, and industrial production. However, in actual applications, gearboxes operate under complex operating conditions for long periods of time and may be affected by factors such as load fluctuations, changes in lubrication conditions, and structural wear, leading to performance degradation or even failure. Therefore, during the running-in bench test, real-time monitoring of the gearbox's operating status and accurate judgment of its operating conditions not only helps to promptly detect potential faults, but also optimizes design parameters, improves product quality, and ensures the stability and reliability of the equipment under actual operating conditions.

[0003] However, gearbox operating condition analysis faces many challenges. On the one hand, the amount of data collected by sensors during the test process is huge, including multiple signals such as vibration, torque, speed, and pressure, accompanied by strong noise interference, which makes the extraction and analysis of effective information complicated. On the other hand, the computing resources of edge devices are limited. Traditional operating condition analysis methods often rely on complex feature extraction and modeling processes, which have bottlenecks in real-time and computational efficiency, making it difficult to meet the dual needs of fast response and high-precision judgment. In addition, the gearbox signal patterns under different operating conditions may vary significantly, resulting in deficiencies in the adaptability and robustness of traditional methods.

[0004] Therefore, there is an urgent need for a comprehensive method that takes into account both real-time processing capabilities and high-precision working condition classification to effectively deal with large-scale, multi-dimensional sensor data and achieve efficient and accurate working condition identification and anomaly detection in edge environments. Summary of the Invention

[0005] The purpose of the present invention is to provide a gearbox operating condition classification method and system to address the shortcomings of existing methods in terms of real-time performance, feature extraction capability and adaptability in an environment with limited edge computing resources.

[0006] In order to achieve the above object, the present invention provides a gearbox operating condition classification method, comprising the following steps:

[0007] The edge server encodes and compresses the collected multimodal sensor signals of the gearbox and uploads them to the cloud server;

[0008] Performing data decoding on the cloud server to obtain the multimodal sensor signal;

[0009] Deploying a teacher model on the cloud server, and performing working condition classification training on the teacher model based on the multimodal sensor signal;

[0010] Deploying a student model on the edge server, performing working condition classification training on the student model based on the multimodal sensor signal, and using the teacher model to guide the training of the student model based on a knowledge distillation mechanism;

[0011] The trained student model is used to predict and classify the working conditions of the gearbox.

[0012] Optionally, the multimodal sensor signal is sequentially differentially encoded and Huffman encoded before being uploaded to the cloud server.

[0013] Optionally, the multimodal sensor signal includes a vibration signal and a non-vibration signal, and performing working condition classification training on the teacher model based on the multimodal sensor signal includes:

[0014] Extracting the global time-frequency feature and / or the first time series feature of the vibration signal and the second time series feature of the non-vibration signal;

[0015] All extracted features are fused to obtain the first fused feature;

[0016] The first classification layer is used to classify the working conditions of the first fusion features and output the corresponding probability distribution.

[0017] Optionally, the vibration signal is subjected to wavelet denoising and short-time Fourier transform processing to generate a time-frequency graph, and then feature extraction is performed on the time-frequency graph based on a ViT network to output the global time-frequency feature.

[0018] Optionally, the vibration signal is subjected to wavelet denoising and then input into an LSTM neural network. The input signal is encoded into a hidden state sequence time step by time by the LSTM neural network, and then the weight of each time step is calculated by an attention module, and the weighted sum is performed to obtain the first time series feature.

[0019] Optionally, the non-vibration signal is input into an LSTM neural network to output the second time series feature.

[0020] Optionally, a cross attention mechanism is used to fuse all the extracted features to obtain a first fused feature.

[0021] Optionally, the multimodal sensor signal includes a vibration signal and a non-vibration signal, and performing working condition classification training on the student model based on the multimodal sensor signal includes:

[0022] The vibration signal is subjected to wavelet denoising and then input into an LSTM neural network to output a third time series feature;

[0023] Inputting the non-vibration signal into the LSTM neural network and outputting a fourth time series feature;

[0024] All the extracted features are fused to obtain the second fused feature;

[0025] The second classification layer is used to classify the second fusion features into working conditions and output the corresponding probability distribution.

[0026] Optionally, when the teacher model is used to guide the training of the student model based on the knowledge distillation mechanism, the KL divergence is calculated using the probability distribution of the teacher model output after softening by the temperature parameter and the probability distribution output by the student model as the distillation loss, and the parameters of the student model are optimized based on the distillation loss and the classification loss of the student model.

[0027] Based on the same inventive concept, the present invention also provides a gearbox operating condition classification system, including a signal acquisition module, an edge server, and a cloud server, wherein:

[0028] The signal acquisition module is used to acquire multimodal sensor signals of the gearbox;

[0029] The edge server is deployed with an encoding module and a student model. The encoding module is used to encode and compress the collected multimodal sensor signals and upload them to the cloud server. The edge server performs working condition classification training on the student model based on the multimodal sensor signals. The student model is used to predict and classify the working condition of the gearbox.

[0030] The cloud server is deployed with a decoding module and a teacher model. The decoding module obtains the multimodal sensor signal through decoding. The cloud server performs working condition classification training on the teacher model based on the multimodal sensor signal, and uses the teacher model to guide the training of the student model based on the knowledge distillation mechanism.

[0031] The gearbox operating condition classification method and system provided by the present invention have at least one of the following beneficial effects:

[0032] 1) By designing a cloud-edge collaborative computing architecture and combining key technologies such as ViT time-frequency graph feature extraction, LSTM time series feature extraction, cross-attention feature fusion, and knowledge distillation, the collaborative training of the cloud-based teacher model and the edge-based student model is achieved. This fully considers the real-time requirements of the working condition classification task and the limitations of insufficient edge computing resources, achieving high-precision, low-latency, and low-cost gearbox working condition classification, which is suitable for intelligent operation and predictive maintenance of construction machinery equipment;

[0033] 2) The edge student model adopts a lightweight design, capable of real-time inference on collected data and outputting the working condition classification results. At the same time, knowledge distillation technology is used to migrate the knowledge of the cloud-based teacher model to the edge student model to maintain high classification accuracy.

[0034] 3) It combines real-time processing capabilities with high-precision classification, effectively handling large-scale, multi-dimensional sensor data and enabling efficient and accurate identification of operating conditions in edge environments. Furthermore, by introducing intelligent analysis technologies such as deep learning and adaptive signal processing, it reduces computational burden while improving model generalization and fault detection accuracy, providing strong support for gearbox health monitoring and intelligent maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Those skilled in the art will appreciate that the accompanying drawings are provided for a better understanding of the present invention and do not constitute any limitation on the scope of the present invention.

[0036] Figure 1 A flow chart of a method for classifying gearbox operating conditions provided by one embodiment of the present invention;

[0037] Figure 2 A flowchart of data encoding and decoding provided by an embodiment of the present invention;

[0038] Figure 3 A schematic diagram of the structure of a teacher model provided by one embodiment of the present invention;

[0039] Figure 4 A schematic structural diagram of a student model provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, advantages and features of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the drawings are in a very simplified form and use non-precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention. In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, please refer to the accompanying drawings. It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the conditions for the implementation of the present invention. Any modification of the structure, change in the proportional relationship or adjustment of the size, under the condition that the effect produced by the present invention and the purpose that can be achieved are the same or similar, should still fall within the scope of the technical content disclosed by the present invention.

[0041] As used herein, the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. As used herein, the term "or" is generally used in a sense including "and / or" unless the context clearly dictates otherwise.

[0042] Please refer to Figure 1 , this embodiment provides a gearbox working condition classification method, comprising the following steps:

[0043] S1. Use the edge server to encode and compress the multimodal sensor signals collected from the gearbox and upload them to the cloud server.

[0044] S2. Decoding data on the cloud server to obtain the multimodal sensor signal;

[0045] S3. Deploy a teacher model on the cloud server, and perform working condition classification training on the teacher model based on the multimodal sensor signal;

[0046] S4. Deploy a student model on the edge server, perform working condition classification training on the student model based on the multimodal sensor signal, and use the teacher model to guide the training of the student model based on a knowledge distillation mechanism;

[0047] S5. Use the trained student model to predict and classify the working conditions of the gearbox.

[0048] By utilizing the teacher model deployed on the cloud server to guide the training of the student model deployed on the edge server based on the knowledge distillation mechanism, collaborative training of the cloud server and the edge server is achieved. This fully considers the real-time requirements of the working condition classification task and the limitations of insufficient edge computing resources, and achieves high-precision, low-latency and low-cost gearbox working condition classification, which is suitable for intelligent operation and maintenance and predictive maintenance of engineering machinery equipment.

[0049] The gearbox operating condition classification method provided in this embodiment can be used for predictive classification of various simulated operating conditions during running-in bench testing, and can also be used for predictive classification of real operating conditions, which is not limited by the present invention.

[0050] First, execute S1 and use the edge server to encode and compress the collected multimodal sensor signals of the gearbox and upload them to the cloud server.

[0051] In this embodiment, the multimodal sensor signal includes multiple vibration signals and non-vibration signals. The non-vibration signals include but are not limited to pressure, speed and torque signals, which can be collected by configuring different sensors at fixed positions on the gearbox, such as main pressure, gear pressure, forward and reverse pressure and oil output pressure pressure sensors, input / output speed sensors, torque sensors and multiple vibration sensors (including torque converter, planetary gear set, power take-off, output end, etc.).

[0052] In this embodiment, Figure 2 As shown, the multimodal sensor signal is sequentially differentially encoded and Huffman encoded before being uploaded to the cloud server. By using this compression method, the low-frequency trend in the signal can be removed, reducing the amount of transmitted data. Specifically, first, the n-way sensor signal {X i (t)} is differentially encoded to reduce data redundancy. The differential encoding formula for any signal is:

[0053] Δx i (t) = x i (t)-x i (t-1)

[0054] Then the differentially encoded data is encoded using Huffman coding. Let the signal discrete value set X be and the probability distribution be P(X). The average coding length satisfies:

[0055]

[0056] By constructing a Huffman coding tree, the data is converted into a corresponding binary code stream, thereby achieving the purpose of compressing the data. The encoded data can be uploaded to the cloud server in real time via the CAN bus and wireless network.

[0057] Then, S2 is executed to decode the data on the cloud server to obtain the multimodal sensor signal. In this embodiment, after receiving the binary data on the cloud, the binary data is first Huffman decoded. The decoding process is based on a pre-built Huffman coding tree, which gradually matches the binary code stream to the corresponding symbols and recovers the differentially encoded symbol sequence. After recovering the differentially encoded data, differential decoding is used to restore it to the original signal. The specific formula is:

[0058] x i (t) = Δx i (t)+x i (t-1),t>0

[0059] By transmitting the initial value x to the cloud server i (0), the compressed data can be fully restored.

[0060] Then, S3 is executed to deploy a teacher model on the cloud server, and to perform working condition classification training on the teacher model based on the multimodal sensor signal.

[0061] Specifically, the multimodal sensor signal includes a vibration signal and a non-vibration signal, and performing working condition classification training on the teacher model based on the multimodal sensor signal includes:

[0062] Extracting the global time-frequency feature and / or the first time series feature of the vibration signal and the second time series feature of the non-vibration signal;

[0063] All extracted features are fused to obtain the first fused feature;

[0064] The first classification layer is used to classify the working conditions of the first fusion features and output the corresponding probability distribution.

[0065] It should be noted that all the features extracted here include the following three cases:

[0066] 1) extracting the global time-frequency features of the vibration signal and the second time series features of the non-vibration signal;

[0067] 2) extracting a first time series feature of the vibration signal and a second time series feature of the non-vibration signal;

[0068] 3) Extracting the global time-frequency features and the first time series features of the vibration signal and the second time series features of the non-vibration signal.

[0069] In this embodiment, combined with Figure 3 As shown in the figure, the teacher model deployed on the cloud server adopts a multi-branch design, mainly including a ViT-based time-frequency graph feature extraction branch and an LSTM-based time series feature extraction branch. Finally, the features of each channel are integrated through a cross-attention mechanism to achieve accurate classification of the gearbox working condition. After the cloud server receives and decodes the data, it first performs wavelet noise reduction on the x-channel vibration signal. The formula is as follows:

[0070]

[0071] The short-time Fourier transform (STFT) is then used to generate the time-frequency diagram, and the formula is as follows:

[0072]

[0073] Where w(t) is the window function. The generated time-frequency graph reflects the local characteristics of the signal in time and frequency. It is the input data of the first branch ViT network. In the ViT network, the time-frequency graph is divided into non-overlapping small blocks (patches). Assuming the patch size is P×P and the time-frequency graph size is H×W, the total number of patches is:

[0074]

[0075] Each patch(x p ) is flattened and linearly mapped to obtain the embedding vector:

[0076]

[0077] in is the projection matrix, C is the number of channels, and D is the embedding dimension. Next, ViT uses a multi-layer Transformer Encoder to encode the global features of the embedding vector. The core self-attention mechanism calculation formula is:

[0078]

[0079] Where Q = ZW Q ,K=ZW K ,V=ZW V are the transformed query, key and value, W Q 、W K 、W V is the trainable weight matrix, D k is the dimension of the attention head. After processing by the multi-head attention layer and the feedforward network, the global time-frequency feature F1 is finally obtained.

[0080] To capture the timing information in the vibration signal, the second branch of the network uses an LSTM network. After wavelet denoising, the x-channel vibration signal is input into the LSTM neural network. A time-step attention module is connected after the LSTM neural network. The input signal is encoded into a hidden state sequence time-step by time-step through the LSTM neural network. The time-step attention module then calculates the weight of each time step, and the weighted summation is used to obtain the first timing feature.

[0081] Specifically, the time-step attention module calculates the input x(t) at each time step to obtain the attention score e(t). Using the additive attention mechanism, we can get:

[0082]

[0083] Among them, h t is the hidden state or input of the current time step, q is the query vector, W a ,W q ,ba are the learnable weight matrix and bias respectively, and the normalized attention weight is:

[0084]

[0085] Among them, α t is the attention weight at time step t, and finally the first time series feature F2 is weighted to be used for subsequent classification processing:

[0086]

[0087] Similarly, in order to capture the timing information of non-vibration signals such as pressure, speed, and torque, the third branch of the network adopts an LSTM network. After decoding and restoration, the y-path pressure / speed / torque signal is directly input into the LSTM neural network for feature extraction to obtain the second timing feature F3 of the non-vibration signal.

[0088] In order to integrate the global time-frequency feature F1 extracted by the ViT network, the first temporal feature F2 extracted by the LSTM neural network, and the first temporal feature F3, this embodiment uses a cross-attention mechanism to achieve deep feature fusion. The fusion formula is:

[0089]

[0090] The attention weight β i Obtained through the following function calculation:

[0091]

[0092] Function φ(·) is the correlation scoring function automatically learned by the deep network. The fused feature F fused The final working condition classification is achieved through the first classification layer, and the corresponding probability distribution is output.

[0093] Then, S4 is executed to deploy a student model on the edge server, perform working condition classification training on the student model based on the multimodal sensor signal, and use the teacher model to guide the training of the student model based on the knowledge distillation mechanism.

[0094] In this embodiment, combined with Figure 4 As shown, the edge student model is designed with a lightweight structure, suitable for real-time reasoning and situations with limited edge computing resources. Its input is the x-path vibration signal processed by wavelet noise reduction and the original y-path non-vibration signal. The working condition classification training of the student model based on the multimodal sensor signals includes:

[0095] The vibration signal is subjected to wavelet denoising and then input into an LSTM neural network to output a third time series feature;

[0096] Inputting the non-vibration signal into the LSTM neural network and outputting a fourth time series feature;

[0097] All the extracted features are fused to obtain the second fused feature;

[0098] The student model also uses the LSTM network to extract time series information and fuses multiple features as follows:

[0099]

[0100] Among them, F x is the third time series feature, F y is the fourth time series feature, γ i The student model adaptively learns the attention weights.

[0101] It should be emphasized that the student model is trained based on multimodal sensor signals and imitates the output distribution of the teacher model through knowledge distillation mechanisms (such as soft label alignment) to achieve lightweight deployment.

[0102] Preferably, when the teacher model is used to guide the training of the student model based on the knowledge distillation mechanism, the KL divergence is calculated using the probability distribution of the teacher model output softened by the temperature parameter and the probability distribution of the student model output as the distillation loss, and the parameters of the student model are optimized based on the distillation loss and the classification loss of the student model. In knowledge distillation, the core role of the KL divergence as a distillation loss is to measure the difference between the probability distributions output by the teacher model and the student model, and to achieve knowledge transfer by minimizing this difference.

[0103] In this embodiment, the probability distribution p output by the teacher model is used teacher The probability distribution p of the student model output student , calculate KL divergence as the distillation loss:

[0104]

[0105] The total loss function is the weighted sum of the classification loss and distillation loss of the student model:

[0106] L total =αL CE +(1-α)L KD

[0107] Among them, L CE is the cross entropy loss, and α is the adjustment parameter used to balance the two parts of the loss.

[0108] Then, back-propagation optimization can be used to update the parameters of the student model by gradient descent so that the output distribution of the student model approaches the output distribution of the teacher model.

[0109] Finally, S5 is executed to predict and classify the gearbox's operating conditions using the trained student model. The student model adopts a lightweight design, capable of real-time inference on collected data and outputting the operating condition classification results. At the same time, knowledge distillation technology is used to migrate the knowledge of the cloud-based teacher model to the edge-side student model to maintain high classification accuracy. In addition, the system will periodically migrate the updated knowledge of the cloud-based teacher model to the edge-side student model through incremental knowledge distillation to achieve continuous optimization of the student model at the edge.

[0110] Based on the same technical concept, the present invention also provides a gearbox operating condition classification system, including a signal acquisition module, an edge server and a cloud server, wherein:

[0111] The signal acquisition module is used to acquire multimodal sensor signals of the gearbox;

[0112] The edge server is deployed with an encoding module and a student model. The encoding module is used to encode and compress the collected multimodal sensor signals and upload them to the cloud server. The edge server performs working condition classification training on the student model based on the multimodal sensor signals. The student model is used to predict and classify the working condition of the gearbox.

[0113] The cloud server is deployed with a decoding module and a teacher model. The decoding module obtains the multimodal sensor signal through decoding. The cloud server performs working condition classification training on the teacher model based on the multimodal sensor signal, and uses the teacher model to guide the training of the student model based on the knowledge distillation mechanism.

[0114] Since the gearbox operating condition classification system provided by the present invention and the gearbox operating condition classification method described above belong to the same inventive concept, the gearbox operating condition classification system provided by the present invention has all the advantages of the gearbox operating condition classification method described above, so the beneficial effects of the gearbox operating condition classification system provided by the present invention will not be described one by one here.

[0115] The technical concept of the present invention is described below through a specific example.

[0116] Data acquisition is performed using three vibration sensors, two speed sensors, one torque sensor, and eight pressure sensors, all of which are fixed to the run-in table. The vibration sensors have an accuracy of ±0.1% and a measurement range of ±5g. The pressure, speed, and torque sensors have accuracies of ±0.5%, ±1%, and ±0.5%, respectively.

[0117] The collected raw signal is first differentially encoded and then compressed using Huffman coding, achieving a data compression ratio of 1:4. The collected data is then uploaded to a cloud server in real time via the CAN bus combined with a 5G wireless network. During data transmission, the static consistency of the data format is maintained, ensuring the continuity of the real-time data stream. After receiving the data, the cloud server performs Huffman decoding and differential decoding sequentially to accurately recover the original signal. Experimental results demonstrate that data recovery accuracy exceeds 99.5%, and minimal latency during dynamic processing ensures efficient data processing.

[0118] The teacher model adopts a multi-branch design. The first branch: wavelet denoising (using Daubechies wavelet, denoising threshold set to 0.05) is first performed on the three vibration signals, then converted into a time-frequency graph through short-time Fourier transform (STFT, window length 256, overlap rate 50%), and then input into the ViT neural network (visual transformer neural network) for feature extraction; the second branch: wavelet denoising (using Daubechies wavelet, denoising threshold set to 0.05) on the same three vibration signals, directly input into the LSTM (long short-term memory network), and combined with the time step attention module to capture the dynamic characteristics of the signal; the third branch: directly input 11 pressure, speed, torque and other signals into another LSTM network; finally, the features extracted from the three branches are deeply fused through the cross-attention mechanism, and the classification output is realized. The classification accuracy of this method is above 99.7%, which is about 6% higher than that of the traditional single model.

[0119] To accommodate resource constraints on edge devices, the student model employs a lightweight design. Three vibration signals undergo wavelet denoising and are then directly fed into an LSTM network along with 11 pressure, speed, and torque signals. A cross-attention mechanism fuses these signal features to achieve fast and accurate operating condition classification. The edge model is trained using knowledge distillation from the cloud-based model. The student model's response latency is kept below 50ms, meeting real-time monitoring requirements. Furthermore, the model's operating condition classification accuracy exceeds 99%, meeting the demand for accurate classification.

[0120] In summary, the embodiments of the present invention provide a gearbox working condition classification method and system. By designing a cloud-side and edge-side collaborative computing architecture and combining key technologies such as ViT time-frequency graph feature extraction, LSTM time series feature extraction, cross-attention feature fusion and knowledge distillation, the collaborative training of the cloud-side teacher model and the edge-side student model is realized. The real-time requirements of the working condition classification task and the limitations of insufficient edge computing resources are fully considered, and high-precision, low-latency and low-cost gearbox working condition classification is achieved, which is suitable for intelligent operation and maintenance and predictive maintenance of engineering machinery equipment.

Claims

1. A gearbox operating condition classification method, characterized in that: The following steps are involved: The edge server encodes and compresses the collected multimodal sensor signals of the gearbox and uploads them to the cloud server; Performing data decoding on the cloud server to obtain the multimodal sensor signal; Deploying a teacher model on the cloud server, and performing working condition classification training on the teacher model based on the multimodal sensor signal; Deploying a student model on the edge server, performing working condition classification training on the student model based on the multimodal sensor signal, and using the teacher model to guide the training of the student model based on a knowledge distillation mechanism; The trained student model is used to predict and classify the working conditions of the gearbox.

2. The gearbox operating condition classification method according to claim 1, characterized in that: The multimodal sensor signal is sequentially differentially encoded and Huffman encoded before being uploaded to the cloud server.

3. The gearbox operating condition classification method according to claim 1, characterized in that: The multimodal sensor signal includes a vibration signal and a non-vibration signal, and performing working condition classification training on the teacher model based on the multimodal sensor signal includes: Extracting the global time-frequency feature and / or the first time series feature of the vibration signal and the second time series feature of the non-vibration signal; All extracted features are fused to obtain the first fused feature; The first classification layer is used to classify the working conditions of the first fusion features and output the corresponding probability distribution.

4. The gearbox operating condition classification method according to claim 3, characterized in that: The vibration signal is subjected to wavelet denoising and short-time Fourier transform processing to generate a time-frequency graph, and then feature extraction is performed on the time-frequency graph based on a ViT network to output the global time-frequency feature.

5. The gearbox operating condition classification method according to claim 3, characterized in that: The vibration signal is subjected to wavelet denoising and then input into an LSTM neural network. The input signal is encoded into a hidden state sequence time step by time by the LSTM neural network. The weight of each time step is then calculated by an attention module, and the weighted sum is performed to obtain the first time series feature.

6. The gearbox operating condition classification method according to claim 3, characterized in that: The non-vibration signal is input into an LSTM neural network, and the second time series feature is output.

7. The gearbox operating condition classification method according to claim 3, characterized in that: The cross attention mechanism is used to fuse all the extracted features to obtain the first fused feature.

8. The gearbox operating condition classification method according to claim 1, characterized in that: The multimodal sensor signal includes a vibration signal and a non-vibration signal. The working condition classification training of the student model based on the multimodal sensor signal includes: The vibration signal is subjected to wavelet denoising and then input into an LSTM neural network to output a third time series feature; Inputting the non-vibration signal into the LSTM neural network and outputting a fourth time series feature; All the extracted features are fused to obtain the second fused feature; The second classification layer is used to classify the second fusion features into working conditions and output the corresponding probability distribution.

9. The gearbox operating condition classification method according to claim 1, characterized in that: When the teacher model is used to guide the training of the student model based on the knowledge distillation mechanism, the KL divergence is calculated using the probability distribution of the teacher model output softened by the temperature parameter and the probability distribution of the student model output as the distillation loss, and the parameters of the student model are optimized based on the distillation loss and the classification loss of the student model.

10. A gearbox operating condition classification system, characterized in that: It includes signal acquisition module, edge server and cloud server, including: The signal acquisition module is used to acquire multimodal sensor signals of the gearbox; The edge server is deployed with an encoding module and a student model. The encoding module is used to encode and compress the collected multimodal sensor signals and upload them to the cloud server. The edge server performs working condition classification training on the student model based on the multimodal sensor signals. The student model is used to predict and classify the working condition of the gearbox. The cloud server is deployed with a decoding module and a teacher model. The decoding module obtains the multimodal sensor signal through decoding. The cloud server performs working condition classification training on the teacher model based on the multimodal sensor signal, and uses the teacher model to guide the training of the student model based on the knowledge distillation mechanism.

Images