AI fault detection method based on optic neural network processor

Abstract

The invention discloses an AI fault detection method based on an optic neural network processor, and belongs to the technical field of AI fault detection.The AI fault detection method based on the optic neural network processor comprises the following specific steps that 1, data collection is conducted, specifically, data in the operation process of equipment is collected in an omnibearing mode through multiple sensors, and the data in the operation process is obtained; and transmitting the acquired image data to an optic neural network processor. The visual neural network processor learns and analyzes a large amount of complex data, various fine fault features can be accurately identified, the fault detection accuracy is greatly improved, the conditions of missing detection and false detection are reduced, compared with a traditional manual detection and simple threshold detection method, a large amount of data can be rapidly processed, and the detection efficiency is improved. The fault is detected in real time, the fault detection efficiency is greatly improved, the model can adapt to different operating environments and new fault types through continuous incremental training, and the method has very high adaptability and expansibility.

Description

Technical Field

[0001] The present invention belongs to the technical field of AI fault detection, and in particular relates to an AI fault detection method based on a visual neural network processor. Background Art

[0002] In many fields such as industrial production and electronic equipment operation, timely and accurate fault detection is crucial to ensure stable system operation and reduce losses. Traditional fault detection methods mostly rely on manual experience judgment or simple threshold settings. For example, in some manufacturing industries, workers judge whether a product is qualified by observing whether there are defects on the appearance of the product with the naked eye, but this method is inefficient and prone to missed detection and false detection. For some complex fault modes, the threshold-based detection method is often difficult to effectively detect potential faults because it cannot accurately capture the complex characteristics of the data.

[0003] A Visual Neural Network Processor (VNNP) is a hardware processor designed specifically for performing visual computing tasks, especially for deep learning and neural network processing, especially tasks involving computer vision. It can efficiently perform visual tasks such as image recognition, object detection, and image segmentation by imitating the way the human brain processes visual information. However, existing visual neural network processors still have low adaptability in detecting faults. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide an AI fault detection method based on a visual neural network processor.

[0005] The technical solution adopted to solve the above technical problems is: an AI fault detection method based on a visual neural network processor, comprising the following specific steps:

[0006] Step 1: Data collection: Use a variety of sensors to collect data from the equipment during operation, and transmit the collected image data to the visual neural network processor;

[0007] Step 2: Data preprocessing: Preprocess the collected image data to improve image quality and facilitate subsequent feature extraction. For the numerical data collected by other sensors, standardize them to the same order of magnitude.

[0008] Step 3: Build a visual neural network model: Use the convolutional neural network (CNN) architecture to build a visual neural network model suitable for fault detection. Train the model with a large amount of sample data with fault labels and continuously adjust the model parameters so that it can accurately identify different types of fault characteristics.

[0009] Step 4: Fault detection: The preprocessed data is input into the trained visual neural network model, and the model outputs the fault detection results. If a fault is detected, the system immediately issues an alarm and locates the location and cause of the fault through data analysis;

[0010] Step 5: Model optimization: Regularly collect new fault data, perform incremental training on the visual neural network model, and continuously optimize the model's performance to adapt to new fault types and complex operating environments.

[0011] Through the above technical solution, various subtle fault characteristics can be accurately identified, which greatly improves the accuracy of fault detection and reduces missed detection and false detection.

[0012] Furthermore, the sensors in step one include a vibration sensor with a collection frequency of 6-9kHz and a temperature sensor with a collection frequency of 3-5Hz. The camera is set to 1080P60fps, and the data collected by the sensors is transmitted to the data center server inside the factory via Ethernet.

[0013] Through the above technical solution, the collected data can be made more accurate and the recognition success rate can be improved.

[0014] Furthermore, the data preprocessing includes the following specific formula:

[0015] Normalize the image data:

[0016] Assume that the image pixel value range is [x min , x max ], to normalize it to [y min ,y max ], for each pixel value x in the image, the normalized pixel value y can be calculated by the following formula:

[0017]

[0018] Standardization is used for numerical data:

[0019] For a set of data x1, x2, ..., x n , whose mean is The standard deviation is

[0020]

[0021] The calculation formula of the standardized value is:

[0022]

[0023] Through the above technical solution, standardization can make data of different magnitudes comparable, which is convenient for neural network processing.

[0024] Furthermore, the convolutional neural network adopts the following specific formula:

[0025] Convolutional Layer:

[0026] Suppose the input feature map X is n×n in size, the convolution kernel is K, the size is k×k, the step size is s, and the padding is p, then the size calculation formula of the output feature map is:

[0027]

[0028] Pooling layer:

[0029] Assume that the input feature map A is of size n×n, the pooling window size is k×k, and the step size is s. The size of the output feature map B is m×m, and the calculation formula is:

[0030]

[0031] The quantification operation can reduce the amount of data while retaining the main features and preventing overfitting;

[0032] Fully connected layer:

[0033] The input vector of the fully connected layer is z = (z1, z2, ..., z m ), the Softmax function calculation formula is:

[0034]

[0035] where y i is the output vector y=(y1, y2, ..., y m ) is used to represent the probability that the input data belongs to the i-th category.

[0036] Furthermore, the model optimization stage adopts the following specific formula:

[0037] Loss function:

[0038] Assume that there are N training data samples. For the i-th sample, its true label is y (i) (is a one-hot vector, for example [0, 1, 0] indicates belonging to the second category), the model prediction output is (also a probability vector), then the cross entropy loss function L is calculated as:

[0039]

[0040] Where C is the number of categories;

[0041] Gradient descent update parameter formula:

[0042] Assume that the model parameter is θ, the loss function is L(θ), and the learning rate is α, then the parameter update formula is:

[0043]

[0044] in is the gradient of the loss function L(θ) with respect to the parameter θ.

[0045] Furthermore, the model optimization phase uses incremental training, collecting new fault data every month and adding it to the original training set. The model is evaluated every quarter using accuracy, recall and F1 value. If the model accuracy is lower than 95%, it is recalibrated.

[0046] The beneficial effects of the present invention are as follows: the present invention can accurately identify various subtle fault characteristics through the learning and analysis of a large amount of complex data by a visual neural network processor, greatly improving the accuracy of fault detection and reducing missed detections and false detections. Compared with traditional manual detection and simple threshold detection methods, the present invention can quickly process a large amount of data and detect faults in real time, greatly improving the efficiency of fault detection and shortening the troubleshooting time. Through continuous incremental training, the model can adapt to different operating environments and newly emerging fault types, and has strong adaptability and scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0049] like Figure 1 As shown, an AI fault detection method based on a visual neural network processor in this embodiment includes the following specific steps:

[0050] Step 1: Data collection: Use a variety of sensors to collect data from the equipment during operation, and transmit the collected image data to the visual neural network processor;

[0051] Step 2: Data preprocessing: Preprocess the collected image data to improve image quality and facilitate subsequent feature extraction. For the numerical data collected by other sensors, standardize them to the same order of magnitude.

[0052] Step 3: Build a visual neural network model: Use the convolutional neural network (CNN) architecture to build a visual neural network model suitable for fault detection. Train the model with a large amount of sample data with fault labels and continuously adjust the model parameters so that it can accurately identify different types of fault characteristics.

[0053] Step 4: Fault detection: The preprocessed data is input into the trained visual neural network model, and the model outputs the fault detection results. If a fault is detected, the system immediately issues an alarm and locates the location and cause of the fault through data analysis;

[0054] Step 5: Model optimization: Regularly collect new fault data, perform incremental training on the visual neural network model, and continuously optimize the model's performance to adapt to new fault types and complex operating environments.

[0055] The sensors in step 1 include a vibration sensor with a collection frequency of 6-9kHz and a temperature sensor with a collection frequency of 3-5Hz. The camera is set to 1080P60fps, and the data collected by the sensors is transmitted to the data center server inside the factory via Ethernet.

[0056] The data preprocessing includes the following specific formula:

[0057] Normalize the image data:

[0058] Assume that the image pixel value range is [x min , x max ], to normalize it to [y min ,y max ], for each pixel value x in the image, the normalized pixel value y can be calculated by the following formula:

[0059]

[0060] Standardization is used for numerical data:

[0061] For a set of data x1, x2, ..., x n , whose mean is The standard deviation is

[0062]

[0063] The calculation formula of the standardized value is:

[0064]

[0065] Standardization can make data of different magnitudes comparable, making it easier for neural networks to process them.

[0066] The convolutional neural network adopts the following specific formula:

[0067] Convolutional Layer:

[0068] Suppose the input feature map X is n×n in size, the convolution kernel is K, the size is k×k, the step size is s, and the padding is p, then the size calculation formula of the output feature map is:

[0069]

[0070] Pooling layer:

[0071] Assume that the input feature map A is of size n×n, the pooling window size is k×k, and the step size is s. The size of the output feature map B is m×m, and the calculation formula is:

[0072]

[0073] The quantification operation can reduce the amount of data while retaining the main features and preventing overfitting;

[0074] Fully connected layer:

[0075] The input vector of the fully connected layer is z = (z1, z2, ..., z m ), the Softmax function calculation formula is:

[0076]

[0077] where y i is the output vector y=(y1, y2, ..., y m ) is used to represent the probability that the input data belongs to the i-th category.

[0078] Loss function:

[0079] Assume that there are N training data samples. For the i-th sample, its true label is y (i) (is a one-hot vector, for example [0, 1, 0] indicates belonging to the second category), the model prediction output is (also a probability vector), then the cross entropy loss function L is calculated as:

[0080]

[0081] Where C is the number of categories;

[0082] Gradient descent update parameter formula:

[0083] Assume that the model parameter is θ, the loss function is L(θ), and the learning rate is α, then the parameter update formula is:

[0084]

[0085] in is the gradient of the loss function L(θ) with respect to the parameter θ.

[0086] The model optimization phase uses incremental training, collecting new fault data every month and adding it to the original training set. The model is evaluated using accuracy, recall and F1 value every quarter. If the model accuracy is lower than 95%, it is recalibrated.

[0087] Specifically, the working process includes the following specific steps:

[0088] Data collection:

[0089] Sensor selection and layout: For existing industrial robots, in order to detect their faults, high-precision vibration sensors are installed at the joints of the robot arms to accurately capture the vibration changes during the operation of the joints. Temperature sensors are attached to the motor housing to monitor the operating temperature of the motor. At the same time, two high-definition cameras are installed above the robot's working area to capture the overall operating status of the robot from different angles to ensure that the subtle movements and appearance of key parts such as the robot arm and gripper can be captured.

[0090] Data collection frequency setting: Since industrial robots run fast and move frequently, the vibration sensor is set to collect data at a frequency of 8kHz to promptly detect vibration anomalies caused by joint wear, etc. The temperature sensor collection frequency is set to 4Hz to meet the monitoring needs of motor temperature changes. The camera frame rate is set to 60fps to clearly record the robot's movements.

[0091] Data transmission and storage: The data collected by the sensor is transmitted to the data center server inside the factory via Ethernet. The server uses RAID5 disk array technology to ensure the security and integrity of the data. The image data collected by the camera is first cached locally and then transferred to the server for storage in batches at regular intervals.

[0092] Data preprocessing:

[0093] Image data preprocessing:

[0094] Grayscale: Convert the color image captured by the camera into a grayscale image.

[0095] Noise reduction: For the Gaussian noise in the image, a 3×3 Gaussian kernel is selected for filtering. Through convolution operation, the noise in the image is removed, making the image clearer and facilitating subsequent feature extraction.

[0096] Normalization: Normalize the image pixel values ​​to a certain range.

[0097] Numerical data preprocessing:

[0098] Outlier processing: During a certain period of time, a data point that deviates significantly from the mean appears in the data collected by the vibration sensor. By calculating the mean and standard deviation of the data, the point is judged as an outlier using the 3σ principle. The interpolation method based on neighborhood data is used to correct the outlier based on the vibration data of adjacent time points.

[0099] Standardization: Standardize the temperature data collected by the temperature sensor. Assuming the mean and standard deviation of the temperature data over a period of time, for the temperature value at a certain moment, complete the standardization according to the formula to make the data comparable.

[0100] Build a visual neural network model:

[0101] Model architecture design:

[0102] Convolution layer design: The constructed model contains 5 convolution layers. The first two layers use 3×3 convolution kernels, and the last three layers use 5×5 convolution kernels. ReLU activation function is added after each convolution layer to increase the nonlinear expression ability of the model.

[0103] Pooling layer design: Two maximum pooling layers and two average pooling layers are interspersed between the convolutional layers. The maximum pooling layer uses a 2×2 pooling window with a step size of 2 to extract key features in the image; the average pooling layer also uses a 2×2 pooling window with a step size of 2 to smooth the feature map. Through the pooling operation, the resolution of the feature map is reduced, reducing the amount of calculation.

[0104] Fully connected layer design: The feature map after multiple convolutions and pooling is expanded and connected to three fully connected layers. The number of neurons in the fully connected layers is 128, 64, and 10 respectively. The last fully connected layer outputs the probability distribution of 10 different fault types through the Softmax function.

[0105] Model training:

[0106] Dataset division: We collected various types of fault data of industrial robots in the past year, including normal operation status data, with a total of 5,000 sets of samples. We divided them into 3,500 training sets, 1,000 validation sets, and 500 test sets in a ratio of 7:2:1.

[0107] Training parameter settings: Stochastic gradient descent algorithm is used, the learning rate is set to 0.001, and the momentum parameter is set to 0.9. The number of training rounds is set to 40, and the training set is traversed once in each round of training.

[0108] Model evaluation and tuning: During the training process, the accuracy and recall of the model were evaluated using the validation set every 5 rounds of training. When the training reached the 25th round, it was found that the accuracy of the model on the validation set began to decline, and overfitting occurred. Therefore, the early stopping method was used to stop the training, and the model structure was adjusted to reduce the number of neurons in the fully connected layer and retrain the model.

[0109] Fault Detection:

[0110] Data input and prediction: The pre-processed image data and numerical data are organized into the input format required by the model and input into the trained visual neural network model. The model outputs the probability distribution of the robot in 10 different fault types through forward propagation calculation.

[0111] Fault judgment and alarm: The fault judgment threshold is set to 0.8. When the probability of a certain fault type output by the model exceeds 0.8, the robot is judged to have this type of fault, and the system immediately triggers a high-decibel sound alarm. At the same time, a text message is sent to the maintenance personnel through the SMS gateway, informing them of the robot number and the fault type, which is a mechanical arm joint fault.

[0112] Fault location and cause analysis: Use the feature information output by the model and combine it with the pre-established fault feature library. For image data, analyze the abnormal features of the robot joints in the feature map and compare them with the image features in the fault feature library. For numerical data, compare the standardized values ​​of vibration data and temperature data with the data patterns in the fault feature library.

[0113] Model optimization:

[0114] Incremental training: New fault data is collected every month and added to the original training set. The initial learning rate is set to 0.001, and the learning rate is halved after every 10 rounds of incremental training.

[0115] Transfer learning: When a new model of industrial robot is introduced, its working principle and some structures are similar to the original robot. Therefore, the parameters of the fault detection model of the original robot are used as the initialization parameters of the new model. In the new scenario, only the fully connected layer is fine-tuned. Through transfer learning, the new model achieves better fault detection results with less training data and shorter training time.

[0116] Model evaluation and continuous optimization: The model is evaluated quarterly using evaluation indicators such as accuracy, recall, and F1 value. If the model accuracy is lower than 95%, the model structure, data quality and other factors are reviewed. If some image data are found to have labeling errors, re-label them and optimize the data quality; at the same time, try to increase the number of convolutional layers, adjust the model structure, and continuously optimize the model performance.

[0117] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention.

Claims

1. An AI fault detection method based on a visual neural network processor, characterized in that: The specific steps include: Step 1: Data collection: Use a variety of sensors to collect data from the equipment during operation, and transmit the collected image data to the visual neural network processor; Step 2: Data preprocessing: Preprocess the collected image data to improve image quality and facilitate subsequent feature extraction. For the numerical data collected by other sensors, standardize them to the same order of magnitude. Step 3: Build a visual neural network model: Use the convolutional neural network (CNN) architecture to build a visual neural network model suitable for fault detection. Train the model with a large amount of sample data with fault labels and continuously adjust the model parameters so that it can accurately identify different types of fault characteristics. Step 4: Fault detection: The preprocessed data is input into the trained visual neural network model, and the model outputs the fault detection results. If a fault is detected, the system immediately issues an alarm and locates the location and cause of the fault through data analysis. Step 5: Model optimization: Regularly collect new fault data, perform incremental training on the visual neural network model, and continuously optimize the model's performance to adapt to new fault types and complex operating environments.

2. The AI ​​fault detection method based on a visual neural network processor according to claim 1 is characterized in that: The sensors in step 1 include a vibration sensor with a collection frequency of 6-9kHz and a temperature sensor with a collection frequency of 3-5Hz. The camera is set to 1080P60fps, and the data collected by the sensors is transmitted to the data center server inside the factory via Ethernet.

3. The AI ​​fault detection method based on a visual neural network processor according to claim 2 is characterized in that: The data preprocessing includes the following specific formula: Normalize the image data: Assume that the image pixel value range is [x min ,x max ], to normalize it to [y min ,y max ], for each pixel value x in the image, the normalized pixel value y can be calculated by the following formula: Standardization is used for numerical data: For a set of data x1,x2,…,x n , whose mean is The standard deviation is The calculation formula of the standardized value is: Standardization can make data of different magnitudes comparable, making it easier for neural networks to process them.

4. The AI ​​fault detection method based on a visual neural network processor according to claim 3 is characterized in that: The convolutional neural network adopts the following specific formula: Convolutional Layer: Suppose the input feature map X is n×n in size, the convolution kernel is K, the size is k×k, the step size is s, and the padding is p, then the size calculation formula of the output feature map is: Pooling layer: Assume that the input feature map A is of size n×n, the pooling window size is k×k, and the step size is s. The size of the output feature map B is m×m. The calculation formula is: The quantification operation can reduce the amount of data while retaining the main features and preventing overfitting; Fully connected layer: The input vector of the fully connected layer is z = (z1, z2, ..., z m ), the Softmax function calculation formula is: where y i The output vector y = (y1, y2, ..., y m ) is used to represent the probability that the input data belongs to the i-th category.

5. The AI ​​fault detection method based on a visual neural network processor according to claim 4 is characterized in that ,The model optimization stage adopts the following specific formula: Loss function: Assume that there are N training data samples. For the i-th sample, its true label is y (i) (is a one-hot vector, for example [0,1,0] means it belongs to the second category), the model prediction output is (also a probability vector), then the cross entropy loss function L is calculated as: Where C is the number of categories; Gradient descent update parameter formula: Assume that the model parameter is θ, the loss function is L(θ), and the learning rate is α, then the parameter update formula is: in is the gradient of the loss function L(θ) with respect to the parameter θ.

6. The AI ​​fault detection method based on a visual neural network processor according to claim 5 is characterized in that: The model optimization phase uses incremental training, collecting new fault data every month and adding it to the original training set. The model is evaluated using accuracy, recall and F1 value every quarter. If the model accuracy is lower than 95%, it is recalibrated.

Images