The HPHT Hydraulic Cubic Press is the "heart" equipment used for manufacturing synthetic diamonds. Since it operates for long periods under the extreme conditions of ultra-high temperature, high pressure, and frequent cyclic loads, its hydraulic system is prone to various issues, such as seal failure, oil contamination, and valve group jamming.

Traditional diagnostic methods are often inefficient and rely heavily on human experience.

To overcome this challenge, experts proposed using Artificial Intelligence (AI), specifically a technique called the Convolutional Neural Network (CNN), to act as the hydraulic system's "smart doctor."

Here is a simplified explanation of how AI is utilized for fault diagnosis in the HPHT Hydraulic Cubic Press hydraulic system:

Step 1: Collecting "Case Files" (Simulation and Data Acquisition)

Training an AI doctor requires a vast amount of fault data, similar to how a human doctor studies numerous case reports.

1. Identifying the "Illnesses": Researchers categorized the common faults in the HPHT Hydraulic Cubic Press hydraulic system into four major "illness types": hydraulic pump, hydraulic valve, hydraulic cylinder (cylinder body), and hydraulic pipeline faults.

    Specific examples of potential faults include hydraulic pump leakage, insufficient flow, valve core damage or jamming, cylinder seal failure, gas in the cylinder body, or pipe joint leakage.

2. Creating a "Virtual Patient": Instead of damaging a real machine, experts built a highly realistic simulation platform of the hydraulic system using AMESim (or Automation Studio). The simulated cylinder diameter was 800 mm.

3. Deliberately Introducing "Symptoms": On this virtual platform, researchers intentionally introduced various fault factors corresponding to the four types of faults (e.g., pump leakage, valve core damage, cylinder seal failure, pipe joint leakage).

4. Collecting "Vital Signs": After a fault was introduced and the system stabilized, monitoring devices were placed at the oil inlets of six hydraulic cylinders to continuously collect two key operating indicators: pressure curves and velocity curves.

5. Preparing the Dataset: To ensure the data was reliable and repeatable, each experimental condition was collected 10 times. Ultimately, 1,440 pressure and velocity curves were collected to form the AI training sample set. Different fault types have different characteristic impacts on these curves; for instance, pump leakage leads to a pressure drop, while gas in the cylinder causes velocity fluctuations.

Step 2: Training the "Smart Doctor" (Convolutional Neural Network CNN)

The CNN algorithm is a type of deep learning network known for its strong feature extraction capability and adaptability. It can automatically identify and extract key features from raw data.

1. Data Input: The collected pressure and velocity curves (like the machine's "ECG" or "X-ray") were processed into training data suitable for the CNN, such as 25 pixels × 25 pixels grayscale images.

2. Network Structure: The diagnostic model was designed based on the LeNet network and improved. The structure typically includes an input layer, convolutional layers, pooling layers, and fully connected layers.

    Convolutional Layer (The Core): This layer extracts local area features through multi-layer convolution operations. The ReLU activation function is used to enable the network to fit non-linear patterns.

    Pooling Layer: This follows the convolutional layer to reduce the feature map size and computation, retaining the most significant features. Max pooling is specifically used because it is beneficial for capturing extreme point feature information.

    Fully Connected Layer: This integrates the local features into global features.

    Output Layer: Using the Softmax function, the output maps to a probability distribution for the classification task, corresponding to the four hydraulic system fault types.

3. Learning and Optimization: The AI model is trained by minimizing the Multi-class Cross-Entropy Loss Function. By adjusting network weights using gradient descent, the model continuously optimizes its diagnostic performance and reduces prediction errors.

Step 3: Assessing the "Doctor's" Diagnostic Accuracy (Results Validation)

The performance of the AI model was validated using a test set (split from the 1,440 samples, typically 8:2 for training/testing).

• Single Input (Pressure): When only the pressure curve was input, the model's accuracy on the test set eventually converged to 1 (100%). The loss function approached 0 after 20 iterations.

• Single Input (Velocity): When only the velocity curve was input, the model's accuracy on the test set also converged to 1 (100%). The loss function approached 0 after 32 iterations.

• Combined Input (Pressure and Velocity): When pressure and velocity (mixed variables) were input jointly, the model's identification accuracy reached over 95%.

In conclusion: The fault diagnosis model built upon the CNN architecture demonstrates strong feature extraction capabilities and high accuracy in fault type recognition. This validates that the CNN-based model can accurately identify fault types, providing an AI solution for the fault recognition of the HPHT Hydraulic Cubic Press hydraulic system.