A fine-grained integrated fault diagnosis method based on knowledge distillation

Abstract

This invention relates to a fine-grained integrated fault diagnosis method based on knowledge distillation. The purpose of this invention is to address the problem that existing fault diagnosis methods cannot simultaneously achieve high accuracy and high robustness. The process is as follows: 1. Collect gear and bearing data using multiple sampling rates under different operating conditions; annotate the different operating conditions of the collected data; 2. Standardize the collected data to obtain standardized gear data; 3. Segment the data for each group using a sliding time window to obtain data corresponding to each time window after segmentation; 4. Construct an FG-KDNet model; 5. Obtain the trained FG-KDNet model; 6. Obtain the gear or bearing data to be tested, input it into the trained model, and the trained model outputs the fault type. This invention is applicable to the field of fault diagnosis.

Description

Technical Field

[0001] This invention relates to a fine-grained integrated fault diagnosis method based on knowledge distillation. Background Technology

[0002] In the context of Industry 4.0, the level of intelligence and automation has become an important direction for technological development in various industries. As key core components in mechanical equipment, gears and bearings play a vital role in the efficient and stable operation of industrial systems. However, under complex and ever-changing dynamic load environments, these components must withstand extreme conditions such as high pressure and high speed, making them prone to wear, fatigue, and structural damage. This leads to increased equipment failure rates, reduced production efficiency, and poses a threat to the safety and reliability of industrial systems.

[0003] Early fault diagnosis research widely utilized signal processing techniques. Researchers analyzed vibration signals to extract parameters closely related to fault characteristics, enabling the identification of fault types and their operating states. With the development of artificial intelligence and big data technologies, neural network-based fault diagnosis methods have been widely applied in equipment health monitoring and fault identification, providing a novel solution to the feature extraction problem. Compared to traditional methods, deep learning models can uncover potential high-dimensional nonlinear features, reduce reliance on domain knowledge, and improve the automation and intelligence of fault diagnosis. In actual production environments, environmental noise and external disturbances weaken the ability of deep learning models to perceive and extract fault features, resulting in insufficient generalization and adaptability under complex and variable operating conditions, making it difficult to fully cope with the diverse challenges of industrial environments.

[0004] To address the aforementioned issues, this invention proposes a robust fault diagnosis method based on knowledge distillation and fine-grained ensemble learning. By employing multi-scale sampling and combining isomorphic knowledge self-learning, knowledge distillation, and ensemble learning methods, the method enhances the model's generalization ability and practical engineering applicability, providing effective support for equipment health monitoring and intelligent diagnosis in intelligent manufacturing scenarios. Summary of the Invention

[0005] The purpose of this invention is to address the problem that existing fault diagnosis methods cannot simultaneously meet the requirements of high accuracy and high robustness, and to propose a fine-grained integrated fault diagnosis method based on knowledge distillation.

[0006] The specific process of a fine-grained integrated fault diagnosis method based on knowledge distillation is as follows:

[0007] Step 1: Collect gear data using multiple sampling rates under different operating conditions; label the different collected data with operating conditions.

[0008] Bearing data is collected using multiple sampling rates under different operating conditions; the different collected data are labeled with operating conditions.

[0009] Step 2: Standardize the gear data collected in Step 1 to obtain standardized gear data;

[0010] The bearing data collected in step one is standardized to obtain standardized bearing data.

[0011] Step 3: Segment the data for each group using a sliding time window to obtain the data corresponding to each time window after segmentation for each group of data;

[0012] Step 4: Construct the FG-KDNet model;

[0013] Step 5: Train the FG-KDNet model based on the data corresponding to each time window until the loss function of the FG-KDNet model converges, and obtain the trained FG-KDNet model.

[0014] Step 6: Obtain the data of the gear or bearing to be tested. Input the data of the gear or bearing to be tested into the trained FG-KDNet model. The trained FG-KDNet model outputs the fault type.

[0015] The beneficial effects of this invention are as follows:

[0016] Strong feature extraction capability: It utilizes multiple sampling rates to achieve diverse expressions of high-frequency and low-frequency features, thereby enhancing the effectiveness of feature extraction;

[0017] The network fusion effect is good: by using the ensemble learning strategy, the fusion of different network features is enhanced, improving feature utilization and diagnostic accuracy;

[0018] Good noise resistance: ensemble learning cancels out the interference of noisy environments on the network model, resulting in small fluctuations in diagnostic accuracy;

[0019] It has broad application prospects: it can be applied to fields such as equipment fault diagnosis, intelligent manufacturing and management.

[0020] In summary, this invention achieves high-precision and robust classification of fault signals, providing an efficient and reliable technical solution for the field of intelligent diagnostics. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention;

[0022] Figure 2 This is a network structure diagram of the present invention. Represents the ResNet18 model; Represents model A1; Represents model A2; Represents model A3; This represents the bidirectional long short-term memory network model BLSTM; Represents model B1; Represents model B2; Model B3 is represented; Knowledge Distillation represents knowledge distillation. This represents a well-trained student model. ; This represents a well-trained student model. ; Represents the ResNet18 model; This represents the bidirectional long short-term memory network model BLSTM; Self-learning is represented by "self-learning"; Dynamic Weight Fusion represents dynamic weight fusion; Collaborative Backprop represents backpropagation. Represents the loss function; Indicates after learning Data Processing refers to data processing; Diagnosis Result refers to the diagnostic result. This indicates stacked learning. Detailed Implementation

[0023] Specific Implementation Method 1: The specific process of this implementation method for fine-grained integrated fault diagnosis based on knowledge distillation is as follows:

[0024] Step 1: Collect gear data using multiple sampling rates under different operating conditions; label the different collected data with the operating conditions (e.g., operating conditions: 200 RPM, 0.2A current).

[0025] Bearing data is collected using multiple sampling rates under different operating conditions; the different collected data are labeled with the operating conditions (e.g., operating conditions: 200 RPM, 0.2A current).

[0026] Step 2: Standardize the gear data collected in Step 1 to obtain standardized gear data;

[0027] The bearing data collected in step one is standardized to obtain standardized bearing data.

[0028] Step 3: Segment the data for each group using a sliding time window to obtain the data corresponding to each time window after segmentation for each group of data;

[0029] Step 4: Construct the FG-KDNet model;

[0030] Step 5: Train the FG-KDNet model based on the data corresponding to each time window until the loss function of the FG-KDNet model converges, and obtain the trained FG-KDNet model.

[0031] Step 6: Obtain the data of the gear or bearing to be tested. Input the data of the gear or bearing to be tested into the trained FG-KDNet model. The trained FG-KDNet model outputs the fault type.

[0032] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that: in step one, gear data is collected using multiple sampling rates under different operating conditions; the different collected data are labeled with operating conditions (e.g., operating conditions are 200 RPM and 0.2A current); bearing data is collected using multiple sampling rates under different operating conditions; the different collected data are labeled with operating conditions (e.g., operating conditions are 200 RPM and 0.2A current); the specific process is as follows:

[0033] Step 1: Under a fixed gear speed operating environment, gear data is collected using sampling rates of 1s, 2s, and 4s respectively. The gear data is then labeled with the operating environment information. The gear data includes speed, vibration, current, and torque during transmission.

[0034] Under the condition of fixed gear vibration, gear data were collected at sampling rates of 1s, 2s, and 4s, respectively. The gear data were labeled with the working environment, and the gear data included speed, vibration, current, and torque during transmission.

[0035] Under the condition of fixed gear current, gear data were collected at sampling rates of 1s, 2s, and 4s, respectively. The gear data was labeled with the operating environment, and the gear data included speed, vibration, current, and torque during transmission.

[0036] Under the working condition of fixed gear transmission torque, gear data were collected using sampling rates of 1s, 2s, and 4s, respectively. The working condition of the gear data was then labeled. The gear data included speed, vibration, current, and torque during transmission.

[0037] Step 1 and Step 2: Under the working condition of fixed bearing speed, the bearing data is collected using sampling rates of 1s, 2s, and 4s respectively. The bearing data is labeled with the working condition. The bearing data includes speed, vibration, current, and torque during transmission.

[0038] Under the condition of fixed bearing vibration, bearing data were collected at sampling rates of 1s, 2s, and 4s, respectively. The bearing data were labeled with the operating environment. The bearing data included speed, vibration, current, and torque during transmission.

[0039] Under the condition of fixed bearing current, bearing data were collected at sampling rates of 1s, 2s, and 4s, respectively. The bearing data were labeled with the operating environment, and the bearing data included speed, vibration, current, and torque during transmission.

[0040] Under the working condition of fixed bearing transmission torque, bearing data were collected using sampling rates of 1s, 2s, and 4s, respectively. The bearing data was labeled with the working condition, including speed, vibration, current, and transmission torque.

[0041] The other steps and parameters are the same as in Specific Implementation Method 1.

[0042] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that: in step two, the gear data collected in step one is standardized to obtain standardized gear data; the bearing data collected in step one is also standardized to obtain standardized bearing data; the specific process is as follows:

[0043] Step 2.1: Standardize the gear data collected in Step 1 to obtain standardized gear data; the specific process is as follows:

[0044] 1) Standardize the gear speed data collected at a fixed sampling rate of 1 second to obtain standardized gear speed data, which will be used as the first set of data;

[0045] The gear vibration data collected at a fixed speed sampling rate of 1 second is standardized to obtain standardized gear vibration data, which is used as the second set of data.

[0046] The gear current data collected at a fixed speed sampling rate of 1 second is standardized to obtain standardized gear current data, which is used as the third set of data.

[0047] The torque data of gear transmission collected at a fixed speed sampling rate of 1 second is standardized to obtain the standardized torque data of gear transmission, which is used as the fourth set of data.

[0048] The gear speed data collected at a fixed sampling rate of 2 seconds were standardized to obtain standardized gear speed data, which was used as the fifth set of data.

[0049] The gear vibration data collected at a fixed speed sampling rate of 2s were standardized to obtain standardized gear vibration data, which was used as the sixth set of data.

[0050] The gear current data collected at a fixed speed sampling rate of 2s were standardized to obtain standardized gear current data, which was used as the seventh set of data.

[0051] The torque data of gear transmission collected at a fixed speed sampling rate of 2s is standardized to obtain the standardized torque data of gear transmission, which is used as the eighth set of data.

[0052] The gear speed data collected at a fixed sampling rate of 4 seconds were standardized to obtain standardized gear speed data, which was used as the ninth set of data.

[0053] The gear vibration data collected at a fixed speed sampling rate of 4s were standardized to obtain standardized gear vibration data, which was used as the tenth set of data.

[0054] The gear current data collected at a fixed speed sampling rate of 4s were standardized to obtain standardized gear current data, which was used as the eleventh set of data.

[0055] The torque data of gear transmission collected at a fixed speed sampling rate of 4s is standardized to obtain the standardized torque data of gear transmission, which is used as the twelfth group of data.

[0056] 2) The gear speed data collected at a sampling rate of 1 second for fixed vibration is standardized to obtain standardized gear speed data, which is used as the thirteenth set of data.

[0057] The gear vibration data collected at a sampling rate of 1 second for fixed vibration were standardized to obtain standardized gear vibration data, which was used as the fourteenth set of data.

[0058] The gear current data collected at a sampling rate of 1 second under fixed vibration were standardized to obtain standardized gear current data, which was used as the fifteenth set of data.

[0059] The torque data of gear transmission collected at a sampling rate of 1 second under fixed vibration were standardized to obtain the standardized torque data during transmission, which was used as the sixteenth set of data.

[0060] The gear speed data collected at a sampling rate of 2s under fixed vibration were standardized to obtain standardized gear speed data, which was used as the seventeenth set of data.

[0061] The gear vibration data collected at a fixed vibration sampling rate of 2s were standardized to obtain standardized gear vibration data, which was used as the eighteenth set of data.

[0062] The gear current data collected at a sampling rate of 2s under fixed vibration were standardized to obtain standardized gear current data, which was used as the nineteenth set of data.

[0063] The torque data of gear transmission collected at a sampling rate of 2s under fixed vibration were standardized to obtain the standardized torque data during transmission, which was used as the twentieth set of data.

[0064] The gear speed data collected at a sampling rate of 4s under fixed vibration were standardized to obtain standardized gear speed data, which was used as the twenty-first data set.

[0065] The gear vibration data collected at a fixed vibration sampling rate of 4s were standardized to obtain standardized gear vibration data, which was used as the twenty-second set of data.

[0066] The gear current data collected at a sampling rate of 4s under fixed vibration were standardized to obtain standardized gear current data, which was used as the twenty-third set of data.

[0067] The torque data of gear transmission collected at a sampling rate of 4s under fixed vibration were standardized to obtain the standardized torque data of gear transmission, which was used as the twenty-fourth set of data.

[0068] 3) The gear speed data collected at a fixed current sampling rate of 1 second is standardized to obtain standardized gear speed data, which is used as the twenty-fifth set of data.

[0069] The gear vibration data collected at a sampling rate of 1 second with a fixed current is standardized to obtain standardized gear vibration data, which is used as the twenty-sixth set of data.

[0070] The gear current data collected at a fixed current sampling rate of 1 second were standardized to obtain standardized gear current data, which was used as the twenty-seventh set of data.

[0071] The torque data of gear transmission collected at a fixed current sampling rate of 1 second is standardized to obtain the standardized torque data of gear transmission, which is used as the twenty-eighth set of data.

[0072] The gear speed data collected at a fixed current sampling rate of 2s were standardized to obtain standardized gear speed data, which was used as the twenty-ninth set of data.

[0073] The gear vibration data collected at a fixed current sampling rate of 2s were standardized to obtain standardized gear vibration data, which was used as the thirtieth set of data.

[0074] The gear current data collected at a fixed current sampling rate of 2s were standardized to obtain standardized gear current data, which was used as the thirty-first data set.

[0075] The torque data of gear transmission collected at a fixed current sampling rate of 2s were standardized to obtain the standardized torque data of gear transmission, which was used as the thirty-second set of data.

[0076] The gear speed data collected at a fixed current sampling rate of 4s were standardized to obtain standardized gear speed data, which was used as the thirty-third set of data.

[0077] The gear vibration data collected at a fixed current sampling rate of 4s were standardized to obtain standardized gear vibration data, which was used as the thirty-fourth set of data.

[0078] The gear current data collected at a fixed current sampling rate of 4s were standardized to obtain standardized gear current data, which was used as the thirty-fifth set of data.

[0079] The torque data of gear transmission collected at a fixed current sampling rate of 4s was standardized to obtain the standardized torque data of gear transmission, which was used as the thirty-sixth set of data.

[0080] 4) The gear speed data collected at a sampling rate of 1 second for torque data during fixed transmission is standardized to obtain standardized gear speed data, which is used as the thirty-seventh set of data.

[0081] The gear vibration data collected at a sampling rate of 1 second during fixed transmission were standardized to obtain standardized gear vibration data, which was used as the thirty-eighth set of data.

[0082] The gear current data collected at a sampling rate of 1 second during fixed transmission torque were standardized to obtain standardized gear current data, which was used as the thirty-ninth set of data.

[0083] The torque data of gear transmission collected at a sampling rate of 1 second under fixed transmission conditions are standardized to obtain standardized torque data of gear transmission, which is used as the 40th set of data.

[0084] The gear speed data collected at a sampling rate of 2s during fixed transmission torque were standardized to obtain standardized gear speed data, which was used as the forty-first data set.

[0085] The gear vibration data collected at a sampling rate of 2s during fixed transmission torque were standardized to obtain standardized gear vibration data, which was used as the forty-second data set.

[0086] The gear current data collected at a sampling rate of 2s during fixed transmission torque were standardized to obtain standardized gear current data, which was used as the forty-third set of data.

[0087] The torque data of gear transmission collected at a sampling rate of 2s during fixed transmission is standardized to obtain standardized torque data of gear transmission, which is used as the forty-fourth set of data.

[0088] The gear speed data collected at a sampling rate of 4s during fixed transmission torque were standardized to obtain standardized gear speed data, which was used as the forty-fifth set of data.

[0089] The gear vibration data collected at a sampling rate of 4s during fixed transmission torque were standardized to obtain standardized gear vibration data, which was used as the forty-sixth set of data.

[0090] The gear current data collected at a sampling rate of 4s during fixed transmission torque were standardized to obtain standardized gear current data, which was used as the forty-seventh set of data.

[0091] The torque data of gear transmission collected at a sampling rate of 4s during fixed transmission is standardized to obtain standardized torque data of gear transmission, which is used as the forty-eighth set of data.

[0092] Step 2: Standardize the bearing data collected in Step 1 to obtain standardized bearing data; the specific process is as follows:

[0093] 1) Standardize the bearing speed data collected at a fixed sampling rate of 1 second to obtain standardized bearing speed data, which will be used as the first set of data and the forty-ninth set of data.

[0094] The bearing vibration data collected at a fixed speed sampling rate of 1 second were standardized to obtain the standardized bearing vibration data, which was used as the second set of data and the fiftieth set of data.

[0095] The bearing current data collected at a fixed speed sampling rate of 1 second were standardized to obtain the standardized bearing current data, which was used as the third set of data and the fifty-first set of data.

[0096] The torque data of bearing transmission collected at a fixed speed sampling rate of 1 second is standardized to obtain the standardized torque data of bearing transmission, which is used as the fourth set of data and the fifty-second set of data.

[0097] The bearing speed data collected at a fixed sampling rate of 2 seconds was standardized to obtain the standardized bearing speed data, which was used as the fifth set of data and the fifty-third set of data.

[0098] The bearing vibration data collected at a fixed speed sampling rate of 2s were standardized to obtain the standardized bearing vibration data, which was used as the fifth set of data and the fifty-fourth set of data.

[0099] The bearing current data collected at a fixed speed sampling rate of 2s were standardized to obtain the standardized bearing current data, which was used as the fifth set of data and the fifty-fifth set of data.

[0100] The torque data of bearing transmission collected at a fixed speed sampling rate of 2s is standardized to obtain the standardized torque data of bearing transmission, which is used as the fifty-sixth set of data.

[0101] The bearing speed data collected at a fixed sampling rate of 4 seconds was standardized to obtain the standardized bearing speed data, which was used as the fifty-seventh set of data.

[0102] The bearing vibration data collected at a fixed sampling rate of 4 seconds was standardized to obtain the standardized bearing vibration data, which was used as the fifty-eighth set of data.

[0103] The bearing current data collected at a fixed speed sampling rate of 4s were standardized to obtain the standardized bearing current data, which was used as the fifty-ninth set of data.

[0104] The torque data of bearing transmission collected at a fixed speed sampling rate of 4s is standardized to obtain the standardized torque data of bearing transmission, which is used as the sixtieth group of data.

[0105] 2) The bearing velocity data collected at a sampling rate of 1 second for fixed vibration is standardized to obtain standardized bearing velocity data, which is used as the sixty-first data set.

[0106] The bearing vibration data collected at a sampling rate of 1 second for fixed vibration were standardized to obtain the standardized bearing vibration data, which was taken as the sixty-second data set.

[0107] The bearing current data collected at a sampling rate of 1 second for fixed vibration were standardized to obtain standardized bearing current data, which was used as the sixty-third data group.

[0108] The torque data of the bearing transmission collected at a sampling rate of 1 second under fixed vibration was standardized to obtain the standardized torque data of the bearing transmission, which was taken as the sixty-fourth set of data.

[0109] The bearing velocity data collected at a sampling rate of 2s under fixed vibration were standardized to obtain standardized bearing velocity data, which was used as the sixty-fifth set of data.

[0110] The bearing vibration data collected at a sampling rate of 2s under fixed vibration were standardized to obtain the standardized bearing vibration data, which was taken as the sixty-sixth set of data.

[0111] The bearing current data collected at a sampling rate of 2s under fixed vibration were standardized to obtain the standardized bearing current data, which was used as the sixty-seventh set of data.

[0112] The torque data of the bearing transmission collected at a sampling rate of 2s under fixed vibration were standardized to obtain the standardized torque data of the bearing transmission, which was taken as the sixty-eighth set of data.

[0113] The bearing velocity data collected at a sampling rate of 4s under fixed vibration were standardized to obtain standardized bearing velocity data, which was used as the sixty-ninth set of data.

[0114] The bearing vibration data collected at a fixed vibration sampling rate of 4s were standardized to obtain the standardized bearing vibration data, which was used as the seventieth group of data.

[0115] The bearing current data collected at a sampling rate of 4s under fixed vibration were standardized to obtain the standardized bearing current data, which was taken as the seventy-first data set.

[0116] The torque data of the bearing transmission collected at a sampling rate of 4s under fixed vibration was standardized to obtain the standardized torque data of the bearing transmission, which was taken as the seventy-second set of data.

[0117] 3) Standardize the bearing speed data collected at a fixed current sampling rate of 1 second to obtain standardized bearing speed data, which will be used as the seventy-third data set.

[0118] The bearing vibration data collected at a sampling rate of 1 second with a fixed current were standardized to obtain the standardized bearing vibration data, which was taken as the seventy-fourth set of data.

[0119] The bearing current data collected at a fixed current sampling rate of 1 second were standardized to obtain standardized bearing current data, which was used as the seventy-fifth set of data.

[0120] The torque data of bearing transmission collected at a fixed current sampling rate of 1 second is standardized to obtain the standardized torque data of bearing transmission, which is used as the seventy-sixth set of data.

[0121] The bearing speed data collected at a fixed current sampling rate of 2s were standardized to obtain standardized bearing speed data, which was used as the seventy-seventh set of data.

[0122] The bearing vibration data collected at a fixed current sampling rate of 2s were standardized to obtain the standardized bearing vibration data, which was taken as the seventy-eighth set of data.

[0123] The bearing current data collected at a fixed current sampling rate of 2s were standardized to obtain the standardized bearing current data, which was used as the seventy-ninth set of data.

[0124] The torque data of the bearing transmission collected at a fixed current sampling rate of 2s is standardized to obtain the standardized torque data of the bearing transmission, which is used as the eightieth set of data.

[0125] The bearing speed data collected at a fixed current sampling rate of 4s is standardized to obtain standardized bearing speed data, which is used as the eighty-first set of data.

[0126] The bearing vibration data collected at a fixed current sampling rate of 4s were standardized to obtain the standardized bearing vibration data, which was used as the eighty-second data set.

[0127] The bearing current data collected at a fixed current sampling rate of 4s were standardized to obtain standardized bearing current data, which was used as the eighty-third set of data.

[0128] The torque data of the bearing transmission collected at a fixed current sampling rate of 4s was standardized to obtain the standardized torque data of the bearing transmission, which was used as the eighty-fourth set of data.

[0129] 4) Standardize the bearing speed data collected at a sampling rate of 1 second for torque data during fixed transmission to obtain standardized bearing speed data, which will be used as the 85th data set.

[0130] The bearing vibration data collected at a sampling rate of 1 second during fixed transmission torque were standardized to obtain standardized bearing vibration data, which was used as the eighty-sixth set of data.

[0131] The bearing current data collected at a sampling rate of 1 second during fixed transmission torque is standardized to obtain standardized bearing current data, which is used as the eighty-seventh set of data.

[0132] The torque data of the bearing transmission collected at a sampling rate of 1 second during fixed transmission is standardized to obtain the standardized torque data of the bearing transmission, which is used as the eighty-eighth set of data.

[0133] The bearing speed data collected at a sampling rate of 2s during fixed transmission torque were standardized to obtain standardized bearing speed data, which was used as the eighty-ninth set of data.

[0134] The bearing vibration data collected at a sampling rate of 2s during fixed transmission torque were standardized to obtain standardized bearing vibration data, which was used as the ninetieth set of data.

[0135] The bearing current data collected at a sampling rate of 2s during fixed transmission torque were standardized to obtain the standardized bearing current data, which was used as the ninety-first data set.

[0136] The torque data of the bearing transmission collected at a sampling rate of 2s during fixed transmission was standardized to obtain the standardized torque data of the bearing transmission, which was taken as the ninety-second set of data.

[0137] The bearing speed data collected at a sampling rate of 4s during fixed transmission torque were standardized to obtain standardized bearing speed data, which was used as the ninety-third data set.

[0138] The bearing vibration data collected at a sampling rate of 4s during fixed transmission torque were standardized to obtain the standardized bearing vibration data, which was used as the ninety-fourth set of data.

[0139] The bearing current data collected at a sampling rate of 4s during fixed transmission torque were standardized to obtain standardized bearing current data, which was used as the ninety-fifth set of data.

[0140] The torque data of the bearing transmission collected at a sampling rate of 4s during fixed transmission was standardized to obtain the standardized torque data of the bearing transmission, which was taken as the ninety-sixth set of data.

[0141] The standardization process uses Z-score standardization to convert the data into a form with zero mean and unit variance, eliminating bias caused by differences in units.

[0142]

[0143] in, For the collected gear or bearing data (one of speed, vibration, current, or torque data during transmission); Standardized gear or bearing data; and These represent the average and standard deviation of gear or bearing data (one of speed, vibration, current, or torque data during transmission) collected at the same sampling rate under the same working conditions.

[0144] Other steps and parameters are the same as in specific implementation method one or two.

[0145] Specific Implementation Method Four: This implementation method differs from one of Specific Implementation Methods One to Three in that: in step three, each group of data is segmented using a sliding time window to obtain data for each time window after segmentation of each group of data.

[0146] The specific process is as follows:

[0147] Set the time window length to 8 seconds and the step size to 1 second.

[0148] For each set of data, a sliding time window is used to segment the data, resulting in data for each time window after segmentation of each set of data.

[0149] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0150] Specific Implementation Method Five: This implementation method differs from one of Specific Implementation Methods One to Four in that: in step four, the FG-KDNet model is constructed;

[0151] The FG-KDNet model includes model A1, model A2, model A3, model B1, model B2, and model B3;

[0152] Models A1, A2, and A3 are ResNet18 models with different initial parameters;

[0153] Models B1, B2, and B3 are bidirectional long short-term memory (BLSTM) network models with different initial parameters.

[0154] The other steps and parameters are the same as in any of the specific implementation methods one to four.

[0155] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One through Five in that: in step five, the FG-KDNet model is trained based on the data corresponding to each time window until the loss function of the FG-KDNet model converges, thus obtaining the trained FG-KDNet model; the specific process is as follows:

[0156] 1) Input all time windows of the first, second, third, and fourth sets of data with fixed speeds into model A1. The penultimate layer of model A1 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0157] The data from all time windows of the fifth, sixth, seventh, and eighth data sets with fixed speeds are input into model A2. The second-to-last layer of model A2 (the output of the layer preceding the FC layer) outputs the feature matrix. ;

[0158] The data from all time windows of the ninth, tenth, eleventh, and twelfth data sets with fixed speeds are input into model A3. The second-to-last layer of model A3 (the output of the layer preceding the FC layer) outputs the feature matrix. ;

[0159] The feature matrix output from the penultimate layer of model A1 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0160] The feature matrix output from the penultimate layer of model A2 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0161] The feature matrix output from the penultimate layer of model A3 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0162] 2) Input the data from all time windows of the thirteenth, fourteenth, fifteenth, and sixteenth data sets for vibration fixation into model A1. The second-to-last layer of model A1 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0163] Data from all time windows of the seventeenth, eighteenth, nineteenth, and twentieth data sets for vibration fixation are input into model A2. The penultimate layer of model A2 (the output of the layer preceding the FC layer) outputs the feature matrix. ;

[0164] The data from all time windows of the 21st, 22nd, 23rd, and 24th data sets for vibration fixation are input into model A3. The second-to-last layer of model A3 (the output of the layer preceding the FC layer) outputs the feature matrix. ;

[0165] The features output from the penultimate layer of model A1 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0166] The features output from the penultimate layer of model A2 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0167] The features output from the penultimate layer of model A3 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0168] 3) Input the data from all time windows of the 25th, 26th, 27th, and 28th data sets with fixed current into model A1. The second-to-last layer of model A1 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0169] The data from all time windows of the 29th, 30th, 31st, and 32nd data sets with fixed current are input into model A2. The second-to-last layer of model A2 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0170] The data from all time windows of the 33rd, 34th, 35th, and 36th data sets with fixed current are input into model A3. The second-to-last layer of model A3 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0171] The feature matrix output from the penultimate layer of model A1 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0172] The feature matrix output from the penultimate layer of model A2 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0173] The feature matrix output from the penultimate layer of model A3 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0174] 4) Input the data from all time windows of the 37th, 38th, 39th, and 40th data sets (for fixed torque data during transmission) into model A1. The second-to-last layer of model A1 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0175] The data from all time windows of the 41st, 42nd, 43rd, and 44th data sets, which are fixed during transmission torque, are input into model A2. The second-to-last layer of model A2 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0176] The data from all time windows of the 45th, 46th, 47th, and 48th data sets, which are fixed for torque data during transmission, are input into model A3. The second-to-last layer of model A3 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0177] The feature matrix output from the penultimate layer of model A1 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0178] The feature matrix output from the penultimate layer of model A2 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0179] The feature matrix output from the penultimate layer of model A3 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0180] 5) Input the data from all time windows of the 49th, 50th, 51st, and 52nd data sets with fixed speeds into model B1. The second-to-last layer of model B1 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0181] The data from all time windows of the 53rd, 54th, 55th, and 56th data sets with fixed speeds are input into model B2. The second-to-last layer of model B2 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0182] The data from all time windows of the 57th, 58th, 59th, and 60th data sets with fixed speeds are input into model B3. The second-to-last layer of model B3 (the output of the layer preceding the FC layer) outputs the feature matrix. ;

[0183] The feature matrix output from the penultimate layer of model B1 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0184] The features output from the penultimate layer of model B2 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0185] The features output from the penultimate layer of model B3 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0186] 6) Input the data from all time windows of the sixty-first, sixty-second, sixty-third, and sixty-fourth data sets for vibration fixation into model B1. The second-to-last layer of model B1 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0187] The data from all time windows of the 65th, 66th, 67th, and 68th data sets for vibration fixation are input into model B2. The second-to-last layer of model B2 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0188] The data from all time windows of the 69th, 70th, 71st, and 72nd data sets with vibration fixation are input into model B3. The second-to-last layer of model B3 (the output of the layer preceding the FC layer) outputs the feature matrix. ;

[0189] The feature matrix output from the penultimate layer of model B1 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0190] The feature matrix output from the penultimate layer of model B2 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0191] The feature matrix output from the penultimate layer of model B3 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0192] 7) Input the data from all time windows of the 73rd, 74th, 75th, and 76th data sets with fixed current into model B1. The second-to-last layer of model B1 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0193] The data from all time windows of the 77th, 78th, 79th, and 80th data sets with fixed current are input into model B2. The second-to-last layer of model B2 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0194] The data from all time windows of the eighty-first, eighty-second, eighty-third, and eighty-fourth data sets with fixed current are input into model B3. The second-to-last layer of model B3 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0195] The feature matrix output from the penultimate layer of model B1 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0196] The feature matrix output from the penultimate layer of model B2 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0197] The feature matrix output from the penultimate layer of model B3 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0198] 8) Input the data from all time windows of the 85th, 86th, 87th, and 88th data sets (for fixed torque data during transmission) into model B1. The second-to-last layer of model B1 (the output of the layer preceding the FC layer) outputs the feature matrix. ;

[0199] The data from all time windows of the 89th, 90th, 91st, and 92nd data sets, which are fixed for torque data during transmission, are input into model B2. The second-to-last layer of model B2 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0200] The data from all time windows of the 93rd, 94th, 95th, and 96th data sets, which are fixed for torque data during transmission, are input into model B3. The second-to-last layer of model B3 (the output of the layer before the FC layer) outputs the feature matrix. ;

[0201] The feature matrix output from the penultimate layer of model B1 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0202] The feature matrix output from the penultimate layer of model B2 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0203] The feature matrix output from the penultimate layer of model B3 Perform weighted fusion to obtain the weighted fused feature matrix. , ;

[0204] 9) The weighted fusion feature matrix obtained in 1) and the feature matrix after weighted fusion The weighted average is then taken to obtain feature matrix 1;

[0205] The weighted fused feature matrix obtained in 2) and the feature matrix after weighted fusion The weighted average is then taken to obtain feature matrix 2;

[0206] The weighted fused feature matrix obtained in 3) and the feature matrix after weighted fusion The weighted average is then taken to obtain feature matrix 3;

[0207] The weighted fused feature matrix obtained from 4) and the feature matrix after weighted fusion The weighted average is then taken to obtain feature matrix 4;

[0208] The weighted fused feature matrix obtained from 5) and the feature matrix after weighted fusion The weighted average is then taken to obtain feature matrix 5;

[0209] The weighted fused feature matrix obtained from 6) and the feature matrix after weighted fusion The weighted average is then taken to obtain feature matrix 6;

[0210] The weighted fused feature matrix obtained from 7) and the feature matrix after weighted fusion The weighted average is then taken to obtain the characteristic matrix 7;

[0211] The weighted fused feature matrix obtained from 8) and the feature matrix after weighted fusion The weighted average is then taken to obtain the feature matrix 8;

[0212] 10) Input feature matrix 1, feature matrix 2, feature matrix 3, and feature matrix 4 into the last layer of model A1. The last layer of model A1 outputs probability values, and the corresponding model A1 is used as the teacher model.

[0213] feature matrix , , , Input the last layer of model A1, and the last layer of model A1 outputs probability values. Use the corresponding model A1 as the student model.

[0214] The student model is trained using knowledge distillation; the loss function for the teacher and student models is... The student model is trained until the loss function of the teacher and student models converges, resulting in a well-trained student model. (ResNet18), a well-trained student model It retains the expert knowledge of each original model on its specific data distribution, and also absorbs the patterns extracted by other models under different data characteristics, thereby improving the model's generalization ability and overall robustness in multi-source or heterogeneous data scenarios.

[0215] 11) Input feature matrix 5, feature matrix 6, feature matrix 7, and feature matrix 8 into the last layer of model B1. The last layer of model B1 outputs probability values. Use the corresponding model B1 as the teacher model.

[0216] feature matrix , , , Input the last layer of model B1, and the last layer of model B1 outputs probability values. Use the corresponding model B1 as the student model.

[0217] The student model is trained using knowledge distillation; the loss function for the teacher and student models is... The student model is trained until the loss function converges, resulting in a well-trained student model. (BLSTM), a well-trained student model It retains the expert knowledge of each original model on its specific data distribution, and also absorbs the patterns extracted by other models under different data characteristics, thereby improving the model's generalization ability and overall robustness in multi-source or heterogeneous data scenarios.

[0218] 12) By integrating the prediction results of multiple complementary base learning models, a meta-learning model is introduced to improve the generalization ability and stability of the overall model. This can preserve the discriminative information of each base model to the maximum extent during the fusion process, avoid feature loss caused by early information compression, and provide richer and more effective feature representations for the final classification decision.

[0219] feature matrix , , , Input the trained student model (ResNet18), a well-trained student model (ResNet18) Output class probability vector:

[0220]

[0221] in, Indicates the total number of fault categories. This represents the predicted probability that the sample belongs to type 1 fault. This represents the predicted probability that the sample belongs to type 2 fault. Indicates that the sample belongs to the first Predicted probability of type of failure; This represents a well-trained student model. (ResNet18) outputs a class probability vector;

[0222] feature matrix , , , Input the trained student model (BLSTM), a well-trained student model (BLSTM) outputs class probability vectors:

[0223]

[0224] in, Indicates the total number of fault categories. This represents the predicted probability that the sample belongs to type 1 fault. This represents the predicted probability that the sample belongs to type 2 fault. Indicates that the sample belongs to the first Predicted probability of type of failure; This represents a well-trained student model. (BLSTM) outputs a class probability vector;

[0225] 13) Convert the class probability vector and Concatenate the columns to obtain a new high-level feature representation vector:

[0226]

[0227] in, Represent real numbers, This represents the new high-level feature representation vector;

[0228] 14) Represent the new high-level feature vector As a meta-learning model Input, meta-learning model Output predicted value ;

[0229] Meta-learning model It consists of a linear mapping layer and a softmax output layer.

[0230] New high-level feature representation vector As a meta-learning model The input is used to further model the nonlinear relationships between the base learning models. In this invention, It is implemented using a shallow fully connected neural network, which includes a linear mapping layer and a softmax output layer to achieve nonlinear mapping and classification from the fused feature space to the fault category space.

[0231] Predicted value Represented as:

[0232]

[0233] in, Meta-learning model The learnable weight matrix, Meta-learning model The bias term, This represents the Softmax activation function, used to map linear outputs to normalized probability distributions for each category;

[0234] Through the stacking and fusion mechanism, different features contained in the fault signal can be extracted, and the complementary advantages of multiple models in feature extraction and pattern recognition can be adaptively integrated to achieve high-precision and robust intelligent fault diagnosis under complex working conditions.

[0235] 15) Set the loss function of the FG-KDNet model; repeat steps 1) to 14) to train the FG-KDNet model until the loss function of the FG-KDNet model converges, and obtain the trained FG-KDNet model.

[0236] The other steps and parameters are the same as those in any of the specific implementation methods one to five.

[0237] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that the teacher-student model loss function... Represented as:

[0238]

[0239] in, Represents the cross-entropy loss function. Indicates divergence loss. Indicates the balance coefficient; Represents the true value of the student model. This indicates that the student model outputs the predicted value. This represents the true value of the teacher model.

[0240] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0241] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One through Seven in that the loss function of the FG-KDNet model is defined as follows:

[0242]

[0243] in, Indicates the number of samples. Let cross-entropy be the loss function. for Regularization term, The regularization coefficient is used. Indicates the fusion weights; This is the parameter set for the FG-KDNet model; For the FG-KDNet model, the first Predicted values ​​for each sample; For the first The true value of each sample.

[0244] Through this joint gradient optimization process, the model can dynamically adjust the weights of sub-networks during training, enabling collaborative learning and information complementarity among multiple models, thereby improving the accuracy and robustness of the fault diagnosis system.

[0245] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0246] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that: in step six, the gear data or bearing data to be tested is obtained, and the gear data or bearing data to be tested is input into the trained FG-KDNet model, which then outputs the fault type; the specific process is as follows:

[0247] Input the gear or bearing data to be tested into the trained student model. (ResNet18), a well-trained student model (ResNet18) outputs a class probability vector;

[0248] Input the gear or bearing data to be tested into the trained student model. (BLSTM), a well-trained student model (BLSTM) outputs a class probability vector;

[0249] The trained student model The class probability vector output by (ResNet18) and the trained student model The class probability vectors output by (BLSTM) are concatenated column-wise to obtain a new high-level feature representation vector. ;

[0250] New high-level feature representation vector As a meta-learning model Input, meta-learning model Output predicted value .

[0251] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0252] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A fine-grained integrated fault diagnosis method based on knowledge distillation, characterized in that: The specific process of the method is as follows: Step 1: Collect gear data using multiple sampling rates under different operating conditions; label the different collected data with operating conditions. Bearing data is collected using multiple sampling rates under different operating conditions; the different collected data are labeled with operating conditions. Step 2: Standardize the gear data collected in Step 1 to obtain standardized gear data; The bearing data collected in step one is standardized to obtain standardized bearing data. Step 3: Segment the data for each group using a sliding time window to obtain the data corresponding to each time window after segmentation for each group of data; Step 4: Construct the FG-KDNet model; Step 5: Train the FG-KDNet model based on the data corresponding to each time window until the loss function of the FG-KDNet model converges, and obtain the trained FG-KDNet model. Step 6: Obtain the data of the gear or bearing to be tested. Input the data of the gear or bearing to be tested into the trained FG-KDNet model. The trained FG-KDNet model outputs the fault type.

2. The fine-grained integrated fault diagnosis method based on knowledge distillation according to claim 1, characterized in that: In step one, gear data is collected using multiple sampling rates under different operating conditions; and the different collected data are labeled with the operating conditions. Bearing data is collected using multiple sampling rates under different operating conditions; the different collected data are labeled with operating conditions. The specific process is as follows: Step 1: Under a fixed gear speed operating environment, gear data is collected using sampling rates of 1s, 2s, and 4s respectively. The gear data is then labeled with the operating environment information. The gear data includes speed, vibration, current, and torque during transmission. Under the condition of fixed gear vibration, gear data were collected at sampling rates of 1s, 2s, and 4s, respectively. The gear data were labeled with the working environment, and the gear data included speed, vibration, current, and torque during transmission. Under the condition of fixed gear current, gear data were collected at sampling rates of 1s, 2s, and 4s, respectively. The gear data was labeled with the operating environment, and the gear data included speed, vibration, current, and torque during transmission. Under the working condition of fixed gear transmission torque, gear data were collected using sampling rates of 1s, 2s, and 4s, respectively. The working condition of the gear data was then labeled. The gear data included speed, vibration, current, and torque during transmission. Step 1 and Step 2: Under the working condition of fixed bearing speed, the bearing data is collected using sampling rates of 1s, 2s, and 4s respectively. The bearing data is labeled with the working condition. The bearing data includes speed, vibration, current, and torque during transmission. Under the condition of fixed bearing vibration, bearing data were collected at sampling rates of 1s, 2s, and 4s, respectively. The bearing data were labeled with the operating environment. The bearing data included speed, vibration, current, and torque during transmission. Under the condition of fixed bearing current, bearing data were collected at sampling rates of 1s, 2s, and 4s, respectively. The bearing data were labeled with the operating environment, and the bearing data included speed, vibration, current, and torque during transmission. Under the working condition of fixed bearing transmission torque, bearing data were collected using sampling rates of 1s, 2s, and 4s, respectively. The bearing data was labeled with the working condition, including speed, vibration, current, and transmission torque.

3. The fine-grained integrated fault diagnosis method based on knowledge distillation according to claim 2, characterized in that: In step two, the gear data collected in step one is standardized to obtain standardized gear data; the bearing data collected in step one is also standardized to obtain standardized bearing data; the specific process is as follows: Step 2.1: Standardize the gear data collected in Step 1 to obtain standardized gear data; the specific process is as follows: 1) Standardize the gear speed data collected at a fixed sampling rate of 1 second to obtain standardized gear speed data, which will be used as the first set of data; The gear vibration data collected at a fixed speed sampling rate of 1 second is standardized to obtain standardized gear vibration data, which is used as the second set of data. The gear current data collected at a fixed speed sampling rate of 1 second is standardized to obtain standardized gear current data, which is used as the third set of data. The torque data of gear transmission collected at a fixed speed sampling rate of 1 second is standardized to obtain the standardized torque data of gear transmission, which is used as the fourth set of data. The gear speed data collected at a fixed sampling rate of 2 seconds were standardized to obtain standardized gear speed data, which was used as the fifth set of data. The gear vibration data collected at a fixed speed sampling rate of 2s were standardized to obtain standardized gear vibration data, which was used as the sixth set of data. The gear current data collected at a fixed speed sampling rate of 2s were standardized to obtain standardized gear current data, which was used as the seventh set of data. The torque data of gear transmission collected at a fixed speed sampling rate of 2s is standardized to obtain the standardized torque data of gear transmission, which is used as the eighth set of data. The gear speed data collected at a fixed sampling rate of 4 seconds were standardized to obtain standardized gear speed data, which was used as the ninth set of data. The gear vibration data collected at a fixed speed sampling rate of 4s were standardized to obtain standardized gear vibration data, which was used as the tenth set of data. The gear current data collected at a fixed speed sampling rate of 4s were standardized to obtain standardized gear current data, which was used as the eleventh set of data. The torque data of gear transmission collected at a fixed speed sampling rate of 4s is standardized to obtain the standardized torque data of gear transmission, which is used as the twelfth group of data. 2) The gear speed data collected at a sampling rate of 1 second for fixed vibration is standardized to obtain standardized gear speed data, which is used as the thirteenth set of data. The gear vibration data collected at a sampling rate of 1 second for fixed vibration were standardized to obtain standardized gear vibration data, which was used as the fourteenth set of data. The gear current data collected at a sampling rate of 1 second under fixed vibration were standardized to obtain standardized gear current data, which was used as the fifteenth set of data. The torque data of gear transmission collected at a sampling rate of 1 second under fixed vibration were standardized to obtain the standardized torque data during transmission, which was used as the sixteenth set of data. The gear speed data collected at a sampling rate of 2s under fixed vibration were standardized to obtain standardized gear speed data, which was used as the seventeenth set of data. The gear vibration data collected at a fixed vibration sampling rate of 2s were standardized to obtain standardized gear vibration data, which was used as the eighteenth set of data. The gear current data collected at a sampling rate of 2s under fixed vibration were standardized to obtain standardized gear current data, which was used as the nineteenth set of data. The torque data of gear transmission collected at a sampling rate of 2s under fixed vibration were standardized to obtain the standardized torque data during transmission, which was used as the twentieth set of data. The gear speed data collected at a sampling rate of 4s under fixed vibration were standardized to obtain standardized gear speed data, which was used as the twenty-first data set. The gear vibration data collected at a fixed vibration sampling rate of 4s were standardized to obtain standardized gear vibration data, which was used as the twenty-second set of data. The gear current data collected at a sampling rate of 4s under fixed vibration were standardized to obtain standardized gear current data, which was used as the twenty-third set of data. The torque data of gear transmission collected at a sampling rate of 4s under fixed vibration were standardized to obtain the standardized torque data of gear transmission, which was used as the twenty-fourth set of data. 3) The gear speed data collected at a fixed current sampling rate of 1 second is standardized to obtain standardized gear speed data, which is used as the twenty-fifth set of data. The gear vibration data collected at a sampling rate of 1 second with a fixed current is standardized to obtain standardized gear vibration data, which is used as the twenty-sixth set of data. The gear current data collected at a fixed current sampling rate of 1 second were standardized to obtain standardized gear current data, which was used as the twenty-seventh set of data. The torque data of gear transmission collected at a fixed current sampling rate of 1 second is standardized to obtain the standardized torque data of gear transmission, which is used as the twenty-eighth set of data. The gear speed data collected at a fixed current sampling rate of 2s were standardized to obtain standardized gear speed data, which was used as the twenty-ninth set of data. The gear vibration data collected at a fixed current sampling rate of 2s were standardized to obtain standardized gear vibration data, which was used as the thirtieth set of data. The gear current data collected at a fixed current sampling rate of 2s were standardized to obtain standardized gear current data, which was used as the thirty-first data set. The torque data of gear transmission collected at a fixed current sampling rate of 2s were standardized to obtain the standardized torque data of gear transmission, which was used as the thirty-second set of data. The gear speed data collected at a fixed current sampling rate of 4s were standardized to obtain standardized gear speed data, which was used as the thirty-third set of data. The gear vibration data collected at a fixed current sampling rate of 4s were standardized to obtain standardized gear vibration data, which was used as the thirty-fourth set of data. The gear current data collected at a fixed current sampling rate of 4s were standardized to obtain standardized gear current data, which was used as the thirty-fifth set of data. The torque data of gear transmission collected at a fixed current sampling rate of 4s was standardized to obtain the standardized torque data of gear transmission, which was used as the thirty-sixth set of data. 4) The gear speed data collected at a sampling rate of 1 second for torque data during fixed transmission is standardized to obtain standardized gear speed data, which is used as the thirty-seventh set of data. The gear vibration data collected at a sampling rate of 1 second during fixed transmission were standardized to obtain standardized gear vibration data, which was used as the thirty-eighth set of data. The gear current data collected at a sampling rate of 1 second during fixed transmission torque were standardized to obtain standardized gear current data, which was used as the thirty-ninth set of data. The torque data of gear transmission collected at a sampling rate of 1 second under fixed transmission conditions are standardized to obtain standardized torque data of gear transmission, which is used as the 40th set of data. The gear speed data collected at a sampling rate of 2s during fixed transmission torque were standardized to obtain standardized gear speed data, which was used as the forty-first data set. The gear vibration data collected at a sampling rate of 2s during fixed transmission torque were standardized to obtain standardized gear vibration data, which was used as the forty-second data set. The gear current data collected at a sampling rate of 2s during fixed transmission torque were standardized to obtain standardized gear current data, which was used as the forty-third set of data. The torque data of gear transmission collected at a sampling rate of 2s during fixed transmission is standardized to obtain standardized torque data of gear transmission, which is used as the forty-fourth set of data. The gear speed data collected at a sampling rate of 4s during fixed transmission torque were standardized to obtain standardized gear speed data, which was used as the forty-fifth set of data. The gear vibration data collected at a sampling rate of 4s during fixed transmission torque were standardized to obtain standardized gear vibration data, which was used as the forty-sixth set of data. The gear current data collected at a sampling rate of 4s during fixed transmission torque were standardized to obtain standardized gear current data, which was used as the forty-seventh set of data. The torque data of gear transmission collected at a sampling rate of 4s during fixed transmission is standardized to obtain standardized torque data of gear transmission, which is used as the forty-eighth set of data. Step 2: Standardize the bearing data collected in Step 1 to obtain standardized bearing data; the specific process is as follows: 1) Standardize the bearing speed data collected at a fixed sampling rate of 1 second to obtain standardized bearing speed data, which will be used as the first set of data and the forty-ninth set of data. The bearing vibration data collected at a fixed speed sampling rate of 1 second were standardized to obtain the standardized bearing vibration data, which was used as the second set of data and the fiftieth set of data. The bearing current data collected at a fixed speed sampling rate of 1 second were standardized to obtain the standardized bearing current data, which was used as the third set of data and the fifty-first set of data. The torque data of bearing transmission collected at a fixed speed sampling rate of 1 second is standardized to obtain the standardized torque data of bearing transmission, which is used as the fourth set of data and the fifty-second set of data. The bearing speed data collected at a fixed sampling rate of 2 seconds was standardized to obtain the standardized bearing speed data, which was used as the fifth set of data and the fifty-third set of data. The bearing vibration data collected at a fixed speed sampling rate of 2s were standardized to obtain the standardized bearing vibration data, which was used as the fifth set of data and the fifty-fourth set of data. The bearing current data collected at a fixed speed sampling rate of 2s were standardized to obtain the standardized bearing current data, which was used as the fifth set of data and the fifty-fifth set of data. The torque data of bearing transmission collected at a fixed speed sampling rate of 2s is standardized to obtain the standardized torque data of bearing transmission, which is used as the fifty-sixth set of data. The bearing speed data collected at a fixed sampling rate of 4 seconds was standardized to obtain the standardized bearing speed data, which was used as the fifty-seventh set of data. The bearing vibration data collected at a fixed sampling rate of 4 seconds was standardized to obtain the standardized bearing vibration data, which was used as the fifty-eighth set of data. The bearing current data collected at a fixed speed sampling rate of 4s were standardized to obtain the standardized bearing current data, which was used as the fifty-ninth set of data. The torque data of bearing transmission collected at a fixed speed sampling rate of 4s is standardized to obtain the standardized torque data of bearing transmission, which is used as the sixtieth group of data. 2) The bearing velocity data collected at a sampling rate of 1 second for fixed vibration is standardized to obtain standardized bearing velocity data, which is used as the sixty-first data set. The bearing vibration data collected at a sampling rate of 1 second for fixed vibration were standardized to obtain the standardized bearing vibration data, which was taken as the sixty-second data set. The bearing current data collected at a sampling rate of 1 second for fixed vibration were standardized to obtain standardized bearing current data, which was used as the sixty-third data group. The torque data of the bearing transmission collected at a sampling rate of 1 second under fixed vibration was standardized to obtain the standardized torque data of the bearing transmission, which was taken as the sixty-fourth set of data. The bearing velocity data collected at a sampling rate of 2s under fixed vibration were standardized to obtain standardized bearing velocity data, which was used as the sixty-fifth set of data. The bearing vibration data collected at a sampling rate of 2s under fixed vibration were standardized to obtain the standardized bearing vibration data, which was taken as the sixty-sixth set of data. The bearing current data collected at a sampling rate of 2s under fixed vibration were standardized to obtain the standardized bearing current data, which was used as the sixty-seventh set of data. The torque data of the bearing transmission collected at a sampling rate of 2s under fixed vibration were standardized to obtain the standardized torque data of the bearing transmission, which was taken as the sixty-eighth set of data. The bearing velocity data collected at a sampling rate of 4s under fixed vibration were standardized to obtain standardized bearing velocity data, which was used as the sixty-ninth set of data. The bearing vibration data collected at a fixed vibration sampling rate of 4s were standardized to obtain the standardized bearing vibration data, which was used as the seventieth group of data. The bearing current data collected at a sampling rate of 4s under fixed vibration were standardized to obtain the standardized bearing current data, which was taken as the seventy-first data set. The torque data of the bearing transmission collected at a sampling rate of 4s under fixed vibration was standardized to obtain the standardized torque data of the bearing transmission, which was taken as the seventy-second set of data. 3) Standardize the bearing speed data collected at a fixed current sampling rate of 1 second to obtain standardized bearing speed data, which will be used as the seventy-third data set. The bearing vibration data collected at a sampling rate of 1 second with a fixed current were standardized to obtain the standardized bearing vibration data, which was taken as the seventy-fourth set of data. The bearing current data collected at a fixed current sampling rate of 1 second were standardized to obtain standardized bearing current data, which was used as the seventy-fifth set of data. The torque data of bearing transmission collected at a fixed current sampling rate of 1 second is standardized to obtain the standardized torque data of bearing transmission, which is used as the seventy-sixth set of data. The bearing speed data collected at a fixed current sampling rate of 2s were standardized to obtain standardized bearing speed data, which was used as the seventy-seventh set of data. The bearing vibration data collected at a fixed current sampling rate of 2s were standardized to obtain the standardized bearing vibration data, which was taken as the seventy-eighth set of data. The bearing current data collected at a fixed current sampling rate of 2s were standardized to obtain the standardized bearing current data, which was used as the seventy-ninth set of data. The torque data of the bearing transmission collected at a fixed current sampling rate of 2s is standardized to obtain the standardized torque data of the bearing transmission, which is used as the eightieth set of data. The bearing speed data collected at a fixed current sampling rate of 4s is standardized to obtain standardized bearing speed data, which is used as the eighty-first set of data. The bearing vibration data collected at a fixed current sampling rate of 4s were standardized to obtain the standardized bearing vibration data, which was used as the eighty-second data set. The bearing current data collected at a fixed current sampling rate of 4s were standardized to obtain standardized bearing current data, which was used as the eighty-third set of data. The torque data of the bearing transmission collected at a fixed current sampling rate of 4s was standardized to obtain the standardized torque data of the bearing transmission, which was used as the eighty-fourth set of data. 4) Standardize the bearing speed data collected at a sampling rate of 1 second for torque data during fixed transmission to obtain standardized bearing speed data, which will be used as the 85th data set. The bearing vibration data collected at a sampling rate of 1 second during fixed transmission torque were standardized to obtain standardized bearing vibration data, which was used as the eighty-sixth set of data. The bearing current data collected at a sampling rate of 1 second during fixed transmission torque is standardized to obtain standardized bearing current data, which is used as the eighty-seventh set of data. The torque data of the bearing transmission collected at a sampling rate of 1 second during fixed transmission is standardized to obtain the standardized torque data of the bearing transmission, which is used as the eighty-eighth set of data. The bearing speed data collected at a sampling rate of 2s during fixed transmission torque were standardized to obtain standardized bearing speed data, which was used as the eighty-ninth set of data. The bearing vibration data collected at a sampling rate of 2s during fixed transmission torque were standardized to obtain standardized bearing vibration data, which was used as the ninetieth set of data. The bearing current data collected at a sampling rate of 2s during fixed transmission torque were standardized to obtain the standardized bearing current data, which was used as the ninety-first data set. The torque data of the bearing transmission collected at a sampling rate of 2s during fixed transmission was standardized to obtain the standardized torque data of the bearing transmission, which was taken as the ninety-second set of data. The bearing speed data collected at a sampling rate of 4s during fixed transmission torque were standardized to obtain standardized bearing speed data, which was used as the ninety-third data set. The bearing vibration data collected at a sampling rate of 4s during fixed transmission torque were standardized to obtain the standardized bearing vibration data, which was used as the ninety-fourth set of data. The bearing current data collected at a sampling rate of 4s during fixed transmission torque were standardized to obtain standardized bearing current data, which was used as the ninety-fifth set of data. The torque data of the bearing transmission collected at a sampling rate of 4s during fixed transmission was standardized to obtain the standardized torque data of the bearing transmission, which was taken as the ninety-sixth set of data. The standardization process uses Z-score standardization.

4. The fine-grained integrated fault diagnosis method based on knowledge distillation according to claim 3, characterized in that: In step three, each group of data is segmented using a sliding time window to obtain data for each time window after segmentation; the specific process is as follows: Set the time window length to 8 seconds and the step size to 1 second. For each set of data, a sliding time window is used to segment the data, resulting in data for each time window after segmentation of each set of data.

5. The fine-grained integrated fault diagnosis method based on knowledge distillation according to claim 4, characterized in that: In step four, the FG-KDNet model is constructed; The FG-KDNet model includes model A1, model A2, model A3, model B1, model B2, and model B3; Models A1, A2, and A3 are ResNet18 models with different initial parameters; Models B1, B2, and B3 are bidirectional long short-term memory (BLSTM) network models with different initial parameters.

6. The fine-grained integrated fault diagnosis method based on knowledge distillation according to claim 5, characterized in that: In step five, the FG-KDNet model is trained based on the data corresponding to each time window until the loss function of the FG-KDNet model converges, thus obtaining the trained FG-KDNet model; the specific process is as follows: 1) Input all time windows of the first, second, third, and fourth data sets with fixed speeds into model A1. The penultimate layer of model A1 outputs the feature matrix. ; The data from all time windows of the fifth, sixth, seventh, and eighth data sets with fixed speeds are input into model A2. The penultimate layer of model A2 outputs the feature matrix. ; The data from all time windows of the ninth, tenth, eleventh, and twelfth data sets with fixed speeds are input into model A3. The penultimate layer of model A3 outputs the feature matrix. ; The feature matrix output from the penultimate layer of model A1 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The feature matrix output from the penultimate layer of model A2 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The feature matrix output from the penultimate layer of model A3 Perform weighted fusion to obtain the weighted fused feature matrix. , ; 2) Input the data from all time windows of the thirteenth, fourteenth, fifteenth, and sixteenth data sets for vibration fixation into model A1. The penultimate layer of model A1 outputs the feature matrix. ; Data from all time windows of the seventeenth, eighteenth, nineteenth, and twentieth data sets for vibration fixation are input into model A2. The penultimate layer of model A2 outputs the feature matrix. ; The data from all time windows of the 21st, 22nd, 23rd, and 24th data sets for vibration fixation are input into model A3. The penultimate layer of model A3 outputs the feature matrix. ; The features output from the penultimate layer of model A1 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The features output from the penultimate layer of model A2 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The features output from the penultimate layer of model A3 Perform weighted fusion to obtain the weighted fused feature matrix. , ; 3) Input the data from all time windows of the 25th, 26th, 27th, and 28th data sets with fixed current into model A1. The penultimate layer of model A1 outputs the feature matrix. ; The data from all time windows of the 29th, 30th, 31st, and 32nd data sets with fixed current are input into model A2. The penultimate layer of model A2 outputs the feature matrix. ; The data from all time windows of the 33rd, 34th, 35th, and 36th data sets with fixed current are input into model A3. The penultimate layer of model A3 outputs the feature matrix. ; The feature matrix output from the penultimate layer of model A1 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The feature matrix output from the penultimate layer of model A2 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The feature matrix output from the penultimate layer of model A3 Perform weighted fusion to obtain the weighted fused feature matrix. , ; 4) Input the data from all time windows of the 37th, 38th, 39th, and 40th data sets (for fixed torque data during transmission) into model A1. The second-to-last layer of model A1 outputs the feature matrix. ; The data from all time windows of the 41st, 42nd, 43rd, and 44th data sets, which are fixed for torque data during transmission, are input into model A2. The second-to-last layer of model A2 outputs a feature matrix. ; The data from all time windows of the 45th, 46th, 47th, and 48th data sets, which are fixed for torque data during transmission, are input into model A3. The second-to-last layer of model A3 outputs a feature matrix. ; The feature matrix output from the penultimate layer of model A1 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The feature matrix output from the penultimate layer of model A2 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The feature matrix output from the penultimate layer of model A3 Perform weighted fusion to obtain the weighted fused feature matrix. , ; 5) Input the data from all time windows of the 49th, 50th, 51st, and 52nd data sets with fixed speeds into model B1. The second-to-last layer of model B1 outputs the feature matrix. ; Data from all time windows of the 53rd, 54th, 55th, and 56th data sets with fixed speeds are input into model B2. The penultimate layer of model B2 outputs a feature matrix. ; Data from all time windows of the 57th, 58th, 59th, and 60th data sets with fixed speeds are input into model B3. The penultimate layer of model B3 outputs a feature matrix. ; The feature matrix output from the penultimate layer of model B1 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The features output from the penultimate layer of model B2 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The features output from the penultimate layer of model B3 Perform weighted fusion to obtain the weighted fused feature matrix. , ; 6) Input the data from all time windows of the sixty-first, sixty-second, sixty-third, and sixty-fourth data sets for vibration fixation into model B1. The second-to-last layer of model B1 outputs the feature matrix. ; Data from all time windows of the 65th, 66th, 67th, and 68th data sets for vibration fixation are input into model B2. The penultimate layer of model B2 outputs the feature matrix. ; Data from all time windows of the 69th, 70th, 71st, and 72nd data sets for vibration fixation are input into model B3. The penultimate layer of model B3 outputs the feature matrix. ; The feature matrix output from the penultimate layer of model B1 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The feature matrix output from the penultimate layer of model B2 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The feature matrix output from the penultimate layer of model B3 Perform weighted fusion to obtain the weighted fused feature matrix. , ; 7) Input the data from all time windows of the 73rd, 74th, 75th, and 76th data sets with fixed current into model B1. The second-to-last layer of model B1 outputs the feature matrix. ; The data from all time windows of the 77th, 78th, 79th, and 80th data sets with fixed current are input into model B2. The penultimate layer of model B2 outputs the feature matrix. ; The data from all time windows of the eighty-first, eighty-second, eighty-third, and eighty-fourth data sets with fixed current are input into model B3. The penultimate layer of model B3 outputs the feature matrix. ; The feature matrix output from the penultimate layer of model B1 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The feature matrix output from the penultimate layer of model B2 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The feature matrix output from the penultimate layer of model B3 Perform weighted fusion to obtain the weighted fused feature matrix. , ; 8) Input all time windows of the 85th, 86th, 87th, and 88th data sets (for fixed torque data during transmission) into model B1. The second-to-last layer of model B1 outputs the feature matrix. ; The data from all time windows of the 89th, 90th, 91st, and 92nd data sets, which are fixed for torque data during transmission, are input into model B2. The second-to-last layer of model B2 outputs a feature matrix. ; The data from all time windows of the 93rd, 94th, 95th, and 96th data sets, which are fixed for torque data during transmission, are input into model B3. The second-to-last layer of model B3 outputs a feature matrix. ; The feature matrix output from the penultimate layer of model B1 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The feature matrix output from the penultimate layer of model B2 Perform weighted fusion to obtain the weighted fused feature matrix. , ; The feature matrix output from the penultimate layer of model B3 Perform weighted fusion to obtain the weighted fused feature matrix. , ; 9) The weighted fusion feature matrix obtained in 1) and the feature matrix after weighted fusion The weighted average is then taken to obtain feature matrix 1; The weighted fused feature matrix obtained in 2) and the feature matrix after weighted fusion The weighted average is then taken to obtain feature matrix 2; The weighted fused feature matrix obtained in 3) and the feature matrix after weighted fusion The weighted average is then taken to obtain feature matrix 3; The weighted fused feature matrix obtained from 4) and the feature matrix after weighted fusion The weighted average is then taken to obtain feature matrix 4; The weighted fused feature matrix obtained from 5) and the feature matrix after weighted fusion The weighted average is then taken to obtain feature matrix 5; The weighted fused feature matrix obtained from 6) and the feature matrix after weighted fusion The weighted average is then taken to obtain feature matrix 6; The weighted fused feature matrix obtained from 7) and the feature matrix after weighted fusion The weighted average is then taken to obtain the characteristic matrix 7; The weighted fused feature matrix obtained from 8) and the feature matrix after weighted fusion The weighted average is then taken to obtain the feature matrix 8; 10) Input feature matrix 1, feature matrix 2, feature matrix 3, and feature matrix 4 into the last layer of model A1. The last layer of model A1 outputs probability values, and the corresponding model A1 is used as the teacher model. feature matrix , , , Input the last layer of model A1, and the last layer of model A1 outputs probability values. Use the corresponding model A1 as the student model. The student model is trained using knowledge distillation; the loss function for the teacher and student models is... The student model is trained until the loss function of the teacher and student models converges, resulting in a well-trained student model. ; 11) Input feature matrix 5, feature matrix 6, feature matrix 7, and feature matrix 8 into the last layer of model B1. The last layer of model B1 outputs probability values. Use the corresponding model B1 as the teacher model. feature matrix , , , Input the last layer of model B1, and the last layer of model B1 outputs probability values. Use the corresponding model B1 as the student model. The student model is trained using knowledge distillation; the loss function for the teacher and student models is... The student model is trained until the loss function converges, resulting in a well-trained student model. ; 12) The characteristic matrix , , , Input the trained student model well-trained student models Output class probability vector: in, Indicates the total number of fault categories. This represents the predicted probability that the sample belongs to type 1 fault. This represents the predicted probability that the sample belongs to type 2 fault. Indicates that the sample belongs to the first Predicted probability of type of failure; This represents a well-trained student model. Output the class probability vector; feature matrix , , , Input the trained student model well-trained student models Output class probability vector: in, Indicates the total number of fault categories. This represents the predicted probability that the sample belongs to type 1 fault. This represents the predicted probability that the sample belongs to type 2 fault. Indicates that the sample belongs to the first Predicted probability of type of failure; This represents a well-trained student model. Output the class probability vector; 13) Convert the class probability vector and Concatenate the columns to obtain a new feature representation vector: in, Represent real numbers, This represents the new feature representation vector; 14) Transform the new feature representation vector As a meta-learning model Input, meta-learning model Output predicted value ; Meta-learning model It consists of a linear mapping layer and a softmax output layer. Predicted value Represented as: in, Meta-learning model The learnable weight matrix, Meta-learning model The bias term, This represents the Softmax activation function; 15) Set the loss function of the FG-KDNet model; repeat steps 1) to 14) to train the FG-KDNet model until the loss function of the FG-KDNet model converges, and obtain the trained FG-KDNet model.

7. The fine-grained integrated fault diagnosis method based on knowledge distillation according to claim 6, characterized in that: The teacher-student model loss function Represented as: in, Represents the cross-entropy loss function. Indicates divergence loss. Indicates the balance coefficient; Represents the true value of the student model. This indicates that the student model outputs the predicted value. This represents the true value of the teacher model.

8. The fine-grained integrated fault diagnosis method based on knowledge distillation according to claim 7, characterized in that: The loss function of the FG-KDNet model is defined as follows: in, Indicates the number of samples. Let cross-entropy be the loss function. for Regularization term, The regularization coefficient is used. Indicates the fusion weights; This is the parameter set for the FG-KDNet model; For the FG-KDNet model, the first Predicted values ​​for each sample; For the first The true value of each sample.

9. The fine-grained integrated fault diagnosis method based on knowledge distillation according to claim 8, characterized in that: In step six, the data of the gear or bearing to be tested is obtained. This data is then input into the trained FG-KDNet model, which outputs the fault type. The specific process is as follows: Input the gear or bearing data to be tested into the trained student model. well-trained student models Output the class probability vector; Input the gear or bearing data to be tested into the trained student model. well-trained student models Output the class probability vector; The trained student model The output class probability vector and the trained student model The output class probability vectors are concatenated column-wise to obtain a new feature representation vector. ; New feature representation vector As a meta-learning model Input, meta-learning model Output predicted value .

Images