Fault diagnosis method for air tightness of additive manufacturing equipment

Abstract

According to the air tightness fault diagnosis method for the additive manufacturing equipment, data of an equipment cabin are collected through a sensor, leakage detection and positioning are achieved in combination with a deep learning algorithm, the method is suitable for a dynamic production environment, problems can be found in time, the downtime can be shortened, and the approximate direction and position of a leakage fault can be positioned; the stability and safety of the equipment manufacturing process are guaranteed, data support is provided for dynamic optimization of technological parameters, and additive manufacturing quality control is converted from experience driving to data driving.

Description

Technical Field

[0001] The present invention relates to the technical field of additive manufacturing, and in particular to a fault diagnosis method for airtightness of additive manufacturing equipment. Background Art

[0002] Additive manufacturing (3D printing), a revolutionary digital manufacturing technology, has profoundly changed the manufacturing paradigm in industries such as aerospace, medical, and automotive. This technology achieves the integral formation of complex components by accumulating materials layer by layer. Its high precision, high material utilization, and customized production capabilities significantly improve manufacturing efficiency. It is worth noting that core metal additive manufacturing processes such as selective laser melting (SLM) and laser direct manufacturing (LDM) all require maintaining a specific controlled environment within the molding chamber: inert gas protection, vacuum environment, or a closed system for powder materials. This plays a decisive role in preventing material oxidation, precisely controlling thermodynamic conditions, and ensuring the performance of the final component.

[0003] The airtightness of the construction chamber is a core element in maintaining these process conditions, and its integrity directly impacts manufacturing quality and safety. In actual production, chamber leakage or seal failure can lead to multiple risks. In the SLM process, the infiltration of trace oxygen will cause metal powder oxidation, significantly reducing component density and mechanical properties. In the LDM process, a decrease in vacuum directly affects the stability of the melt pool and causes microstructural defects. Furthermore, abnormal fluctuations in environmental parameters will trigger system protection mechanisms, forcing the printing process to be interrupted, resulting in equipment downtime, material waste, and extended production cycles.

[0004] When flammable metal powders or reactive materials are involved, failure of the cabin's airtight seal may trigger a dust explosion or the spread of harmful substances, causing serious safety accidents and posing a serious threat to the health of operators and the ecological environment. Summary of the Invention

[0005] The purpose of the present invention is to provide a fault diagnosis method for the airtightness of additive manufacturing equipment. By collecting data from the equipment cabin through sensors and combining it with deep learning algorithms, leak detection and positioning can be achieved. The method is suitable for dynamic production environments, can detect problems in a timely manner, reduce downtime and locate the approximate direction and position of leakage faults. It not only ensures the stability and safety of the equipment manufacturing process, but also provides data support for the dynamic optimization of process parameters, enabling the transformation of additive manufacturing quality control from experience-driven to data-driven.

[0006] According to a first aspect of the present invention, a method for diagnosing airtightness faults in additive manufacturing equipment is provided, comprising the following steps:

[0007] A pressure sensor, an air intake duct flow meter, an exhaust duct flow meter and an oxygen content sensor are installed in the equipment construction cabin, which are respectively used to detect the cabin pressure, air intake flow, exhaust flow and the oxygen concentration of the mixed gas in the cabin; wherein the pressure sensor and the oxygen content sensor are deployed in groups, and are respectively arranged at the cabin top, left side, right side, front and back positions of the equipment construction cabin, and a first solenoid valve for simulating cabin leakage is arranged at the above five positions at the same time; the above-mentioned air intake duct flow meter and exhaust duct flow are respectively set at the positions of the air intake duct and the exhaust duct close to the cabin end, and a second solenoid valve for simulating pipeline leakage is respectively arranged in the air intake duct and the exhaust duct;

[0008] The simulated air leakage levels were divided into three levels according to the rotation angle of the solenoid valve ball valve. Based on these levels, the first solenoid valve was operated to perform multiple groups of cabin leakage scrubbing tests, and the second solenoid valve was operated to perform multiple groups of pipeline leakage scrubbing tests. During each group of scrubbing tests, data was collected using the pressure sensor, intake pipeline flow meter, exhaust pipeline flow meter, and oxygen content sensor.

[0009] Preprocess the data collected from multiple groups of scrubbing tests to obtain standardized multi-source feature data through noise reduction and normalization.

[0010] Extracting derived feature data based on standardized multi-source feature data, and constructing input feature data based on the multi-source feature data and the derived feature data;

[0011] Based on a data set consisting of input feature data for each fault category and normal state, a CNN-based deep learning network is trained to obtain a CNN-based recognition model.

[0012] The trained recognition model is used to extract input feature data to obtain high-dimensional input features. The high-dimensional input features are mapped to a two-dimensional plane using the T-SNE algorithm to form a visual scatter plot. Based on the visualization results, the model parameters are adjusted (for example, increasing the number of CNN layers or adjusting the number of LSTM hidden units). Repeated training and model optimization are performed until the expected fault clustering visualization results are obtained, and a gas leakage fault recognition model is obtained. Different colors in the visual scatter plot represent different fault categories.

[0013] During actual use of the equipment, data on cabin pressure, intake flow, exhaust flow, and oxygen concentration of the mixed gas in the cabin are collected in real time. After preprocessing and inputting feature data into the construction, the data is input into the leakage fault identification model, and the identified fault type is output as at least one of cabin roof leakage, left side leakage, right side leakage, front leakage, back leakage, and pipeline leakage; and

[0014] According to the identified fault type, the leakage position and propagation direction are located in combination with the spatial distribution of sensors.

[0015] As an optional embodiment, the simulated leakage level is divided into three levels according to the rotation angle of the ball valve of the solenoid valve, including: dividing the leakage level into three levels according to the rotation angle of the ball valve of the first solenoid valve and the second solenoid valve being 30°, 60°, and 90° respectively;

[0016] Therefore, during the air scrubbing test, at least three groups of air scrubbing tests are performed at each air leakage level according to the three levels. The operation of each group is as follows:

[0017] Operate the first solenoid valve or the second solenoid valve at a position to simulate a gas leakage failure at the position where the first solenoid valve or the second solenoid valve is located, and sample the relevant data of the mixed gas in the cabin within a preset time period through the pressure sensor, the intake pipe flow meter, the exhaust pipe flow meter and the oxygen content sensor, wherein each group of gas washing tests simulates a leakage at one position.

[0018] In an optional embodiment, extracting derived feature data based on standardized multi-source feature data includes:

[0019] Calculate the pressure change rate ΔP / Δt based on the sliding window period Δt, where ΔP represents the pressure change value of one sliding window period;

[0020] The standard deviation of oxygen concentration within the sliding window period Δt, representing its fluctuation amplitude; and

[0021] Gas flow deviation ΔQ, ΔQ represents the ratio of the actual gas flow deviation to the set flow.

[0022] In an optional embodiment, the CNN-based deep learning network adopts a composite model of CNN-LSTM, including an input layer, a convolutional layer, an LSTM layer, and a fully connected layer;

[0023] Input layer: used to receive input feature data of sliding window period Δt. Feature data dimension = number of sensors × number of features × time step;

[0024] Convolutional layer: 2 layers of 1D convolution, convolution kernel size 3, number of channels 32→64, extracting local spatial features;

[0025] LSTM layer: 1 bidirectional LSTM layer with 128 hidden units to capture temporal dependency features;

[0026] Fully connected layer: Dropout is used to suppress overfitting, and the Softmax function is used for activation in the output layer, corresponding to 6 types of fault AF and normal state, a total of 7 categories.

[0027] Model training is carried out on this basis. The training process includes:

[0028] The dataset including normal samples and various fault samples is divided into training set, validation set and test set according to the ratio of 7:2:1;

[0029] The CNN-based deep learning network was trained using the training set, with the Adam optimizer learning rate set to 0.001 and the decay rate set to 1e-5. The training was iterated for 50-100 rounds, and the training was stopped when the validation set accuracy did not improve for 5 consecutive rounds. This resulted in a CNN-based recognition model.

[0030] Based on the test set, the accuracy and F1 score are used for evaluation.

[0031] In an optional embodiment, locating the leakage location based on the identified fault type and in combination with the spatial distribution of sensors includes:

[0032] Based on the data collected during actual use of the equipment, including cabin pressure, intake flow, exhaust flow, and oxygen concentration of the mixed gas in the cabin, a sliding window (taking 10 seconds as an example) is used for detection. The standard deviation of the data within the window is calculated and the abnormal time series is marked accordingly to obtain the abnormal starting time point of each sensor.

[0033] Arrange the abnormal starting time points of all sensors in ascending time order to form an event sequence E, E={e1,e2,...,e n}, where ei = (sensor ID, abnormal starting time point t i );

[0034] The sensor S with the earliest abnormal start time m Determined to be closest to the leak source.

[0035] Furthermore, the sensor S with the earliest abnormal start time is used. m To accurately determine the direction of the leakage propagation path, the following process is specifically included:

[0036] With sensor S m As the center, calculate the abnormal starting time of other sensors relative to t m Gradient t m Indicates sensor S m The abnormal starting time point, d j For sensor S j to S m The Euclidean distance, t j Indicates sensor S j The abnormal starting time point;

[0037] like The gas diffusion source direction is determined by conforming to a preset gas diffusion physical model, for example, the gas diffusion physical model selects a horizontal diffusion model or a spherical diffusion model according to the airflow characteristics in the cabin.

[0038] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below, as long as such concepts are not mutually inconsistent, can be considered part of the inventive subject matter of this disclosure. In addition, all combinations of the claimed subject matter are considered part of the inventive subject matter of this disclosure.

[0039] The foregoing and other aspects, embodiments, and features of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the present invention, such as features and / or beneficial effects of the exemplary embodiments, will become apparent from the following description or through practice of specific embodiments according to the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings are not intended to be drawn to scale. In the accompanying drawings, each identical or nearly identical component shown in various figures may be represented by the same reference numeral. For clarity, not every component is labeled in every figure. Embodiments of various aspects of the present invention will now be described by way of example and with reference to the accompanying drawings.

[0041] Figure 1 4 is a flow chart of a method for diagnosing airtightness faults of additive manufacturing equipment according to an embodiment of the present invention.

[0042] Figure 2 4 is a schematic diagram of an offline processing stage of a fault diagnosis method for airtightness of additive manufacturing equipment according to an embodiment of the present invention.

[0043] Figure 3 4 is a schematic diagram of an online diagnosis processing stage of a fault diagnosis method for airtightness of additive manufacturing equipment according to an embodiment of the present invention.

[0044] Figure 4 2 is a schematic diagram of a model training process of a fault diagnosis method for air tightness of additive manufacturing equipment according to an embodiment of the present invention.

[0045] Figure 5 It is a schematic diagram of the layout design of the cabin, sensor and solenoid valve in a fault test according to a specific example of the present invention. DETAILED DESCRIPTION

[0046] In order to better understand the technical content of the present invention, specific embodiments are given below in conjunction with the accompanying drawings.

[0047] Aspects of the present application are described in the disclosure with reference to the accompanying drawings, in which a number of illustrative embodiments are shown. The embodiments of this disclosure need not necessarily include all aspects of the present application. It should be understood that various concepts and embodiments introduced above and those described in more detail below can be implemented in any of numerous ways, as the disclosed concepts and embodiments are not limited to any one implementation. Additionally, some aspects of the present disclosure can be used independently of any other aspects of the present disclosure, or in any appropriate combination.

[0048] {Example 1}

[0049] In conjunction Figures 1 to 4 As shown, the fault diagnosis method for air tightness of an additive manufacturing device according to the embodiments of the present application comprises:

[0050] S101, a pressure sensor, an inlet pipeline flow meter, an exhaust pipeline flow meter, and an oxygen content sensor are installed in the device build cabin, for detecting the cabin pressure, the inlet flow, the exhaust flow, and the oxygen concentration of the mixed gas in the cabin, respectively; wherein the pressure sensor and the oxygen content sensor are arranged in groups and arranged at the top, left side, right side, front, and back of the device build cabin, and a first electromagnetic valve for simulating cabin air leakage is arranged at the same time at the above five positions; the inlet pipeline flow meter and the exhaust pipeline flow meter are arranged at the positions close to the cabin end of the inlet pipeline and the exhaust pipeline, respectively, and a second electromagnetic valve for simulating pipeline air leakage is arranged in the inlet pipeline and the exhaust pipeline, respectively;

[0051] S102, according to the ball valve rotation angle of the electromagnetic valve, the simulated air leakage level is divided into three levels, and the first electromagnetic valve is operated to perform a plurality of cabin leakage gas washing tests, and the second electromagnetic valve is operated to perform a plurality of pipeline leakage gas washing tests; during each group of gas washing tests, data is collected by the pressure sensor, the inlet pipeline flow meter, the exhaust pipeline flow meter, and the oxygen content sensor, respectively;

[0052] S103, the data collected in the plurality of gas washing tests is preprocessed, and standardized multi-source feature data is obtained through noise reduction and normalization processing;

[0053] S104, derivative feature data is extracted based on the standardized multi-source feature data, and input feature data is constructed according to the multi-source feature data and the derivative feature data;

[0054] S105, based on the data set composed of the input feature data of each fault category and the normal state, a CNN-based deep learning network is used for training to obtain a CNN-based recognition model;

[0055] S106. Using the trained recognition model to extract input feature data to obtain high-dimensional input features, the high-dimensional input features are mapped to a two-dimensional plane using the T-SNE algorithm to form a visual scatter plot. Model parameters are adjusted based on the visualization results, and training is repeated to optimize the model until the expected fault clustering visualization results are obtained, thereby obtaining a gas leakage fault recognition model, where different colors in the visual scatter plot represent different fault categories.

[0056] S107. During actual use of the equipment, real-time data on cabin pressure, intake flow, exhaust flow, and oxygen concentration of the mixed gas in the cabin are collected, pre-processed, and input into a leakage fault identification model after inputting feature data into the model. The model then outputs the identified fault type, which is at least one of cabin roof leakage, left side leakage, right side leakage, front leakage, back leakage, and pipe leakage.

[0057] S108. Locate the leakage position and propagation direction based on the identified fault type and the spatial distribution of sensors.

[0058] As an optional implementation, in step S102, the simulated leakage level is divided into three levels according to the rotation angle of the ball valve of the solenoid valve, including: dividing the leakage level into three levels according to the rotation angle of the ball valve of the first solenoid valve and the second solenoid valve being 30°, 60°, and 90° respectively.

[0059] Therefore, during the air scrubbing test, at least three groups of air scrubbing tests are performed at each air leakage level according to the three levels. The operation of each group is as follows:

[0060] Operate the first solenoid valve or the second solenoid valve at a position to simulate a gas leakage failure at the position where the first solenoid valve or the second solenoid valve is located, and use the pressure sensor, the intake pipe flow meter, the exhaust pipe flow meter and the oxygen content sensor to sample the relevant data of the mixed gas in the cabin within a preset time period (taking 60 minutes as an example), where each group of gas washing tests simulates a leakage at one position.

[0061] In an optional embodiment, in S103, extracting derived feature data based on the standardized multi-source feature data includes:

[0062] Calculate the pressure change rate ΔP / Δt based on the sliding window period Δt, where ΔP represents the pressure change value of one sliding window period;

[0063] The standard deviation of oxygen concentration within the sliding window period Δt, representing its fluctuation amplitude; and

[0064] Gas flow deviation ΔQ, ΔQ represents the ratio of the actual gas flow deviation to the set flow.

[0065] In an optional embodiment, the collected historical data can be further pre-processed and cleaned, for example, using wavelet transform (db4 wavelet, 3-layer decomposition) for noise reduction to filter high-frequency noise (such as sensor circuit interference). Furthermore, outliers (such as sudden sensor jumps) can be identified based on the 3σ principle and replaced or eliminated through linear interpolation.

[0066] The aforementioned data normalization standard processing uses min-max normalization and maps it to the [0,1] interval to achieve standardization of data of different dimensions.

[0067] In an optional embodiment, the aforementioned CNN-based deep learning network adopts a CNN-LSTM composite model, including an input layer, a convolutional layer, an LSTM layer, and a fully connected layer;

[0068] Input layer: used to receive input feature data of sliding window period Δt. Feature data dimension = number of sensors × number of features × time step;

[0069] Convolutional layer: 2 layers of 1D convolution, convolution kernel size 3, number of channels 32→64, extracting local spatial features;

[0070] LSTM layer: 1 bidirectional LSTM layer with 128 hidden units to capture temporal dependency features;

[0071] Fully connected layer: Dropout is used to suppress overfitting, and the Softmax function is used for activation in the output layer, corresponding to 6 types of fault AF and normal state, a total of 7 categories.

[0072] On this basis, the model training process includes:

[0073] The dataset including normal samples and various fault samples is divided into training set, validation set and test set according to the ratio of 7:2:1;

[0074] The CNN-based deep learning network was trained using the training set, with the Adam optimizer learning rate set to 0.001 and the decay rate set to 1e-5. The training was iterated for 50-100 rounds, and the training was stopped when the validation set accuracy did not improve for 5 consecutive rounds. This resulted in a CNN-based recognition model.

[0075] Based on the test set, the accuracy and F1 score are used for evaluation.

[0076] In an optional embodiment, the visualization results of the T-SNE algorithm are used to adjust model parameters, including:

[0077] Increase the number of CNN layers and / or adjust the number of LSTM hidden units.

[0078] In an optional embodiment, the leak location is located in combination with the sensor spatial distribution according to the identified fault type, including:

[0079] According to the data of the cabin pressure, the air inlet flow, the exhaust flow and the oxygen concentration of the mixed gas in the cabin actually collected during actual use of the device, the abnormal time sequence is marked by sliding window detection, the standard deviation of the data in the window is calculated, and the abnormal starting time point of each sensor is obtained;

[0080] The abnormal starting time points of all sensors are arranged in ascending order of time to form an event sequence E, E={e1, e2,..., en}, where e n i i =(sensor ID, abnormal starting time point t i );

[0081] The sensor S m with the earliest abnormal starting time is determined as the closest to the leak source.

[0082] Wherein, taking the standard deviation σ of the sensor data in the window as the reference, when σ exceeds the 3σ threshold value of the historical normal state, the time point t is marked as the abnormal starting time point of the sensor.

[0083] Wherein, the sensor S m with the earliest abnormal starting time is used as the reference to determine the leak propagation path direction, including:

[0084] Taking the sensor S m as the center, the gradient of the abnormal starting time point of other sensors relative to t m is calculated t m represents the abnormal starting time point of the sensor S m , j is the Euclidean distance from the sensor S j to S m , and t j represents the abnormal starting time point of the sensor S j ;

[0085] If the gas diffusion physical model is consistent with the preset gas diffusion physical model, for example, the gas diffusion physical model selects an advection diffusion model or a spherical diffusion model according to the air flow characteristics in the cabin, the gas leak source direction is confirmed.

[0086] {Example 2}

[0087] In this embodiment, the combination of Figure 1 , 2 , 3 and Figure 5As shown, pressure and gas sensors are first installed within the build chamber to continuously monitor the pressure and gas composition. These sensors provide real-time data, capturing dynamic changes in the chamber's internal environment. To improve positioning accuracy, multiple sensors are placed at different locations within the chamber to ensure coverage of all potential leak areas.

[0088] Six categories of fault types are defined:

[0089]

[0090] Multiple groups of simulated leakage tests were designed according to the leakage levels. According to the leakage levels corresponding to the ball valve rotation angles of the solenoid valve (30°, 60°, 90°), the simulated leakage levels were divided into three levels: slight leakage, moderate leakage and severe leakage. Based on this, the first solenoid valve was operated to carry out multiple groups of cabin leakage washing tests, and the second solenoid valve was operated to carry out multiple groups of pipeline leakage washing tests.

[0091] During each set of air scrubbing tests, data are collected respectively through the pressure sensor, the intake pipe flow meter, the exhaust pipe flow meter and the oxygen content sensor.

[0092] The simulated air leakage test is as follows:

[0093]

[0094] During the operation phase, the data collected by the current sensor is input into the trained model in real time, and the model identifies the abnormal fault category output, indicating the existence of leakage.

[0095] At the same time, after detecting data anomalies, the direction and propagation path of the fault leakage point are further analyzed based on the spatial positioning of the sensor to locate its approximate position, so as to discover problems in time, reduce downtime, and ensure the normal operation and safety of additive manufacturing equipment.

[0096] While the present invention has been disclosed above with reference to preferred embodiments, this is not intended to limit the present invention. Persons skilled in the art will readily appreciate that various modifications and variations can be made without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for diagnosing the airtightness of additive manufacturing equipment, characterized in that: include: A pressure sensor, an air intake duct flow meter, an exhaust duct flow meter and an oxygen content sensor are installed in the equipment construction cabin, which are respectively used to detect the cabin pressure, air intake flow, exhaust flow and the oxygen concentration of the mixed gas in the cabin; wherein the pressure sensor and the oxygen content sensor are deployed in groups, and are respectively arranged at the cabin top, left side, right side, front and back positions of the equipment construction cabin, and a first solenoid valve for simulating cabin leakage is arranged at the above five positions at the same time; the above-mentioned air intake duct flow meter and exhaust duct flow are respectively set at the positions of the air intake duct and the exhaust duct close to the cabin end, and a second solenoid valve for simulating pipeline leakage is respectively arranged in the air intake duct and the exhaust duct; The simulated air leakage levels were divided into three levels according to the rotation angle of the solenoid valve ball valve. Based on these levels, the first solenoid valve was operated to perform multiple groups of cabin leakage scrubbing tests, and the second solenoid valve was operated to perform multiple groups of pipeline leakage scrubbing tests. During each group of scrubbing tests, data was collected using the pressure sensor, intake pipeline flow meter, exhaust pipeline flow meter, and oxygen content sensor. Preprocess the data collected from multiple groups of scrubbing tests to obtain standardized multi-source feature data through noise reduction and normalization. Extracting derived feature data based on standardized multi-source feature data, and constructing input feature data based on the multi-source feature data and the derived feature data; Based on a data set consisting of input feature data for each fault category and normal state, a CNN-based deep learning network is trained to obtain a CNN-based recognition model. The trained recognition model is used to extract input feature data to obtain high-dimensional input features. The high-dimensional input features are mapped to a two-dimensional plane using the T-SNE algorithm to form a visual scatter plot. The model parameters are adjusted based on the visualization results, and training is repeated to optimize the model until the expected fault clustering visualization results are obtained. This results in a gas leakage fault recognition model, where different colors in the visual scatter plot represent different fault categories. During actual use of the equipment, data on cabin pressure, intake flow, exhaust flow, and oxygen concentration of the mixed gas in the cabin are collected in real time. After preprocessing and inputting feature data into the construction, the data is input into the leakage fault identification model, and the identified fault type is output as at least one of cabin roof leakage, left side leakage, right side leakage, front leakage, back leakage, and pipeline leakage; and According to the identified fault type, the leakage position and propagation direction are located in combination with the spatial distribution of sensors.

2. The method for diagnosing airtightness of additive manufacturing equipment according to claim 1, characterized in that: According to the rotation angle of the ball valve of the solenoid valve, the simulated leakage level is divided into three levels, including: The air leakage levels are divided into three levels according to the rotation angles of the ball valves of the first solenoid valve and the second solenoid valve being 30°, 60°, and 90° respectively.

3. The method for diagnosing the airtightness of additive manufacturing equipment according to claim 1, characterized in that: During the air scrubbing test, at least three groups of air scrubbing tests are performed at each air leakage level according to the three levels. The operation of each group is as follows: Operate the first solenoid valve or the second solenoid valve at a position to simulate a gas leakage failure at the position where the first solenoid valve or the second solenoid valve is located, and sample the relevant data of the mixed gas in the cabin within a preset time period through the pressure sensor, the intake pipe flow meter, the exhaust pipe flow meter and the oxygen content sensor, wherein each group of gas washing tests simulates a leakage at one position.

4. The method for diagnosing the airtightness of additive manufacturing equipment according to claim 1, characterized in that: The method of extracting derived feature data based on standardized multi-source feature data includes: The pressure change rate ΔP / Δt is calculated based on the sliding window period Δt, where ΔP represents the pressure change value in one sliding window period (taking 10s as an example); The standard deviation of oxygen concentration within the sliding window period Δt, representing its fluctuation amplitude; and Gas flow deviation ΔQ, ΔQ represents the ratio of the actual gas flow deviation to the set flow.

5. The method for diagnosing airtightness faults of additive manufacturing equipment according to claim 1, characterized in that: The CNN-based deep learning network adopts a CNN-LSTM composite model, including an input layer, a convolutional layer, an LSTM layer, and a fully connected layer; Input layer: used to receive input feature data of sliding window period Δt. Feature data dimension = number of sensors × number of features × time step; Convolutional layer: 2 layers of 1D convolution, convolution kernel size 3, number of channels 32→64, extracting local spatial features; LSTM layer: 1 bidirectional LSTM layer with 128 hidden units to capture temporal dependency features; Fully connected layer: Dropout is used to suppress overfitting, and the Softmax function is used for activation in the output layer, corresponding to 6 types of fault AF and normal state, a total of 7 categories.

6. The method for diagnosing airtightness faults of additive manufacturing equipment according to claim 1, characterized in that: The training process of the model includes: The dataset including normal samples and various fault samples is divided into training set, validation set and test set according to the ratio of 7:2:1; The CNN-based deep learning network was trained using the training set, with the Adam optimizer learning rate set to 0.001 and the decay rate set to 1e-5. The training was iterated for 50-100 rounds, and the training was stopped when the validation set accuracy did not improve for 5 consecutive rounds. This resulted in a CNN-based recognition model. Based on the test set, the accuracy and F1 score are used for evaluation.

7. The method for diagnosing airtightness faults of additive manufacturing equipment according to claim 1, characterized in that: The visualization results of the T-SNE algorithm adjust the model parameters, including: Increase the number of CNN layers and / or adjust the number of LSTM hidden units.

8. The method for diagnosing airtightness faults of additive manufacturing equipment according to any one of claims 1 to 7, characterized in that: The method of locating the leakage location based on the identified fault type and the spatial distribution of sensors includes: Based on the data collected during actual use of the equipment, including cabin pressure, intake flow, exhaust flow, and oxygen concentration of the mixed gas in the cabin, a sliding window detection is performed to calculate the standard deviation σ of the sensor data within the window. When σ exceeds the 3σ threshold of the historical normal state, the time point t is marked as the sensor abnormality start time point. The abnormality start time point of each sensor is thus obtained. Arrange the abnormal starting time points of all sensors in ascending time order to form an event sequence E, E={e1,e2,...,e n }, where e i =(sensor ID, abnormal start time t i ); The sensor S with the earliest abnormal start time m Determined to be closest to the leak source.

9. The method for diagnosing airtightness faults of additive manufacturing equipment according to claim 8, characterized in that: The sensor S with the earliest abnormal start time m To determine the direction of the leakage propagation path, including: With sensor S m As the center, calculate the abnormal starting time of other sensors relative to t m Gradient t m Indicates sensor S m The abnormal starting time point, d j For sensor S j to S m The Euclidean distance, t j Indicates sensor S j The abnormal starting time point; like If it conforms to the preset gas diffusion physical model, the direction of the gas leakage source is confirmed.

10. The method for diagnosing airtightness faults of additive manufacturing equipment according to claim 9, characterized in that: The gas diffusion physical model selects a horizontal diffusion model or a spherical diffusion model according to the airflow characteristics in the cabin to determine the direction of the gas leakage source.

Images