A Precision Machine Tool Linear Axis Feed Motion Creep Measurement Device and Intelligent Recognition Method

Abstract

A precision machine tool linear axis feed motion crawling measurement device and intelligent recognition method belong to the field of precision measurement technology. The precision machine tool linear axis feed motion crawling measurement device includes a mounting plate, a precision adjustment platform, a connecting plate, a sensor fastening fixture, a protective baffle, a detection block, and a sensor mounted on the sensor fastening fixture. The sensor is connected to the machine tool column via the mounting plate, and the remaining components are mounted on the mounting plate. First, the machine tool feed axis drive chain is defined to identify the final actuator that reflects the crawling phenomenon. Then, a high-precision capacitive displacement sensor is selected as the core acquisition hardware to monitor the displacement of the final actuator. The sensor is installed in the fixture. Next, the acquired data is analyzed and filtered. Based on a deep learning algorithm, the data is trained to achieve the key discrimination of the crawling phenomenon. This invention ensures stable measurement even under actual machine tool processing conditions. The combination of traditional moving average filtering signal processing and deep learning algorithms results in high accuracy in crawling phenomenon discrimination.

Description

Technical Field

[0001] This invention belongs to the field of precision measurement technology, and relates to a measurement device and intelligent identification method for crawling motion of linear axis feed motion of precision machine tools. Specifically, it relates to a measurement device and a method for identifying low-speed crawling phenomena during the feed motion of linear axis of precision machine tools. Background Technology

[0002] Precision machine tools play a vital role in enhancing my country's industrial competitiveness. However, linear axes on precision machine tools are prone to creeping during low-speed precision feed motions, which directly affects the machining accuracy and surface quality of the workpiece.

[0003] Chinese invention patent CN 110193753A proposes monitoring crawling phenomena using servo encoder signals built into the CNC system. This method calculates slide displacement by measuring motor rotation angle. While it offers fast response and easy integration, it essentially only reflects the motion of the drive end and cannot capture nonlinear errors such as gaps and deformations in the mechanical transmission chain (e.g., lead screws, guide rails, couplings) under actual loads and disturbances. Therefore, it struggles to accurately characterize the crawling features of the final actuator. Chinese invention patent CN 2043906U proposes a combination of grating sensors and silicon photodiodes to indirectly obtain displacement signals by detecting changes in light transmittance. However, its optical principle is susceptible to interference from ambient light, oil, and vibration, resulting in poor signal stability. Furthermore, the sensor is magnetically mounted on the guide rail surface, making stable clamping difficult due to chips, coolant, and vibration during processing. This prevents continuous and stable online monitoring and limits its application to no-load or offline monitoring scenarios.

[0004] The aforementioned studies all used indirect measurements and could not directly reflect the crawling phenomenon. Therefore, this invention proposes a method that can monitor motion status in real-time, with high precision, and directly in a real-world processing environment, and effectively identify crawling characteristics. Summary of the Invention

[0005] To achieve the above objectives, this invention provides a precision machine tool linear axis feed motion crawling measurement device and intelligent recognition method. The invention first clarifies the transmission chain of the precision machine tool linear axis feed motion, identifying the final actuator that will react to crawling phenomena. Then, a high-precision capacitive displacement sensor is selected as the core acquisition hardware to monitor the displacement of the final actuator. The sensor is mounted in a dedicated fixture, which is designed to possess key functions. The acquired data is then analyzed and filtered, and trained using a deep learning algorithm, ultimately enabling the key discrimination of crawling phenomena.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A high-precision measuring device for the linear axis feed motion crawling of a precision machine tool includes a mounting plate 5, a precision adjustment platform 6, a connecting plate 7, a sensor fastening fixture 8, a protective baffle 9, a detection block 10, and a sensor mounted on the sensor fastening fixture 8. To ensure sufficient stability during measurement, the device is connected to the column 1 of the precision machine tool via the mounting plate 5, with the remaining components mounted on the mounting plate 5. Specifically: Mounting plate 5 has a rectangular plate structure, which is used to vertically mount on column 1. All other components are mounted on mounting plate 5 to ensure stability during the measurement process.

[0007] Considering the small detection range of the selected capacitive displacement sensor (the sensor can only collect data when the distance between the sensor probe and the detected surface is less than 200 micrometers), to accurately set the initial gap between the sensor probe and the detected object to be less than 200 micrometers, the sensor fastening clamp 8 and the fastened sensor need to be able to make minor adjustments in the vertical direction. Therefore, a precision adjustment platform 6 is used for minor adjustments in the vertical direction. This precision adjustment platform 6 is equipped with an adjustment knob, and rotating the knob converts the rotational motion into linear motion through a high-precision threaded pair. The precision adjustment platform 6 is connected to the mounting plate 5. A connecting plate 7 is designed, with an overall rectangular plate structure. The right side of the connecting plate 7 is connected to the precision adjustment platform 6, and the left side is designed with two vertical bosses arranged in parallel. The distance between the two bosses is the width of the sensor fastening clamp 8, serving as a guide groove. The sensor fastening clamp 8 is placed in the guide groove and connected to the connecting plate 7 to ensure that the sensor fastening clamp 8 is placed vertically, thereby ensuring that the sensor is placed vertically. Rotating the adjustment knob drives the entire connecting plate 7 and the sensor fastening clamp 8, along with the fastened sensor, to make precise vertical displacement, thereby achieving precise adjustment of the initial gap.

[0008] A sensor clamping fixture 8 is designed to secure the sensor. Its structure consists of two arms, an inner arm 8-1 and an outer arm 8-2. The inner arm 8-1 is longer than the outer arm 8-2. The inner arm 8-1 connects to the guide groove of the connecting plate 7. Each arm has a semi-cylindrical groove perpendicularly designed on its inner side to hold the cylindrical sensor probe. An optical hole is designed on the outer arm 8-2, and a threaded hole is designed at the corresponding position on the inner arm 8-1. A bolt passes through the optical hole on the outer arm 8-2 and is screwed into the threaded hole on the inner arm 8-1. During bolt tightening, the two arms gradually close together, generating clamping force to fix the sensor probe. A threaded hole is designed at the top of the sensor clamping fixture 8, which cooperates with the protective baffle 9 for fine adjustments.

[0009] To prevent sudden changes in the linear axis feed speed of a precision machine tool from causing the horizontal portion of the detection block 10 to collide with the sensor probe, a protective baffle 9 is designed to prevent the sensor probe from being struck first during a sudden change in linear axis feed speed. Specifically: The protective baffle 9 is a rectangular plate structure, placed horizontally above and connected to the connecting plate 7. A through-hole, slightly larger than the sensor probe diameter, is designed in the center of the protective baffle 9, allowing the sensor probe to pass through and measure the horizontal portion of the detection block 10. During sensor installation, manual adjustment ensures the sensor probe is slightly below the upper surface of the protective baffle 9. Therefore, in the event of a sudden change in the linear axis feed speed causing an impact, the horizontal portion of the detection block 10 above the protective baffle 9 first impacts the upper surface of the protective baffle 9, preventing direct impact on the sensor probe below, thus effectively protecting the sensor probe.

[0010] Furthermore, to accurately set and maintain the aforementioned protective positional relationship, a fine-tuning mechanism is provided on the protective baffle 9. The right end of the protective baffle 9 has two holes, one on the left and one on the right. The left hole is a through-hole, which engages with the threaded hole above the sensor fastening fixture 8 for minor adjustments. A bolt is screwed through the through-hole into the threaded hole above the sensor fastening fixture 8. During tightening, the main body of the sensor fastening fixture 8 and the fastened sensor probe are pulled upwards relative to the protective baffle 9, bringing the sensor probe closer to the upper surface of the protective baffle 9. The right hole is a through-hole, which is screwed into the threaded hole above the protective baffle 9. The bolt is screwed in to a depth exceeding the depth of the threaded hole, with the end of the bolt pressing against the upper surface of the sensor fastening fixture 8. During tightening, the main body of the sensor fastening fixture 8 and the fastened sensor probe are pressed downwards relative to the protective baffle 8, moving the sensor probe away from the upper surface of the protective baffle 9. By alternately adjusting the left and right bolts, ensure that the upper surface of the protective baffle 9 is always above the sensor probe to protect the sensor probe.

[0011] Analyzing the linear axis feed motion transmission chain of a precision machine tool, the transmission motion is determined by a motor rotating to drive a lead screw, which in turn drives a lead screw nut to move along the linear guide rail 4, thereby moving the slide 2 connected to the lead screw nut and slider 3. It is clear that the final component reflecting the crawling phenomenon is the slide 2. Therefore, the displacement of the slide 2 needs to be detected during the testing process. Considering the limitations of the testing space and the requirement for sufficient flatness of the tested surface, a detection block 10 is designed as the actual testing object. This L-shaped plate is divided into a vertical section and a horizontal section. The vertical section is connected to the side of the slide 2 and is located on both sides of the connecting plate 7 with the sensor fastening fixture 8, maintaining a certain distance from the connecting plate 7. The horizontal section is located above the protective baffle 9. When the linear axis moves vertically downwards, the horizontal section of the detection block 10 gradually approaches the protective baffle 9. When the distance between the detection block 10 and the sensor probe fastened by the sensor fastening fixture 8 below the upper surface of the protective baffle 9 is less than the sensor's detection range, the sensor detects the displacement of the horizontal section of the detection block 10.

[0012] A method for intelligent identification of crawling measurement of linear axis feed motion in precision machine tools, based on the aforementioned measuring device, includes the following steps: The first step is to install the measuring device. Once the measuring device is fully installed, measurements can begin. Specifically: Step 1.1: Connect the mounting plate 5 to the column 1 to obtain the fixed mounting plate 5; Step 1.2: For the fixed mounting plate 5, connect the precision adjustment platform 6 to the mounting plate 5 to obtain the fixed precision adjustment platform 6; Step 1.3: For the fixed precision adjustment platform 6, connect the right side of the connecting plate 7 to the precision adjustment platform 6 to obtain the fixed connecting plate 7.

[0013] Step 1.4: For the fixed connecting plate 7, the sensor fastening clamp 8 is initially fixed to the left guide groove of the connecting plate 7, but not locked, so as to facilitate subsequent adjustments, and the sensor fastening clamp 8 is initially fixed.

[0014] Step 1.5: For the fixed connecting plate 7, connect the protective baffle 9 to the upper end of the connecting plate 7 to obtain the fixed protective baffle 9.

[0015] Step 1.6: Use bolts to pass through the light hole of the outer clamping arm 8-2 and screw them into the threaded hole of the inner clamping arm 8-1 to gradually bring the two clamping arms together to generate clamping force to fix the sensor probe.

[0016] Step 1.7: Use bolts to alternately adjust the light hole arranged on the upper right end of the protective baffle 9 and the threaded hole designed on the upper part of the sensor fastening clamp 8, as well as the threaded hole on the right side of the light hole, so that the sensor probe is located below the upper surface of the protective baffle 9. Then, finally fix the sensor fastening clamp 8, which was initially fixed in step 1.4, to the connecting plate 7.

[0017] Step 1.8: Connect the vertical part of the detection block 10 to the side of the slide 2, and the horizontal part is located above the protective baffle 9, thus obtaining the measuring device after the whole assembly is completed.

[0018] The second step involves adjusting the parameters of the precision machine tool control panel after the measuring device has been fully installed in the first step to obtain measurement data; specifically: Step 2.1: Change the linear axis feed motion parameters on the precision machine tool control panel to make the linear axis feed downwards. The slide 2 on the linear axis also feeds downwards, thereby driving the detection block 10 connected to the side of the slide 2 to feed downwards.

[0019] Furthermore, since a capacitive displacement sensor with a range of 200 micrometers is selected, the sensor can collect data when the distance between the sensor probe and the measured surface is less than 200 micrometers. When the horizontal part of the visual inspection block 10 is close to the sensor probe but the sensor does not show a reading, the linear axis movement stops, and the adjustment knob of the precision adjustment platform 6 is adjusted. When the adjustment is made so that the sensor shows a reading, the adjustment knob is stopped.

[0020] Step 2.2: When the sensor displays a reading, readjust the precision machine tool control panel to the parameters to be measured, start the linear axis to begin feeding, and control the sensor to start collecting data, finally obtaining the measurement data.

[0021] The third step, based on the measurement data obtained in the second step, involves preliminary data processing and training a deep learning algorithm to accurately identify crawling phenomena. The identification process is as follows: Figure 5 The signal processing and crawling recognition flowchart is shown below. Specifically: Step 3.1: Differentiate the original displacement data obtained in step 2 to obtain the original velocity data. Then, apply a moving average filter to the velocity data to obtain a noise-removed velocity sequence. The filtered signal at the current moment is: It is given by the following formula: in, It is the first one The filtered output value The length of the sliding window. The first input signal Each sample value. This processing can significantly improve the signal-to-noise ratio, providing stable and smooth input data for subsequent feature recognition.

[0022] Step 3.2: Construct a temporal classification model based on a Convolutional Neural Network (CNN). This model takes the velocity sequence filtered by the moving average in Step 3.1 as input. The velocity sequence is divided into several velocity segments with a fixed window length (256 columns), each segment being an independent sample. To generate a label for each sample, the velocity difference between the same velocity segment under crawling and non-crawling conditions is calculated. This velocity difference is compared to a set difference threshold (0.1 μm / s). If the difference is greater than the threshold, crawling is considered to exist; if it is less than the threshold, non-crawling is considered to exist. The true state ("crawling" or "normal") at each sampling time is determined and recorded as 0 or 1. Then, the label for each velocity segment is determined by the state of the majority of sampling points within that segment. If the majority of points are crawling, the label is 1 (crawling); otherwise, it is 0 (normal). All labeled samples are randomly divided into a training set and a validation set in an 8:2 ratio, with the validation set used to monitor the model's generalization performance. Design a one-dimensional convolutional neural network for temporal classification, consisting of three convolutional modules, a global average pooling layer, and a fully connected classification layer connected in series. Each convolutional module includes a convolutional layer, a batch normalization layer, a linear rectified activation layer, and a max pooling layer. Specifically: In velocity sequences, crawling is characterized by sudden velocity changes within a short time window, with complete crawling events and their intervals typically lasting from tens of milliseconds to several seconds. Convolutional layers are the core component of convolutional modules, their function being to extract local features of the velocity sequence using learnable convolutional kernels. As the specific feature extraction operator within a convolutional layer, a single convolutional layer, having only one set of learnable kernels, is limited by its fixed receptive field (i.e., the time span covered by the kernel on the input velocity sequence), and can only extract limited details within that receptive field, making it difficult to simultaneously capture both local transient features and global event trends. Therefore, this network employs three convolutional modules, progressively increasing the receptive field layer by layer. This allows shallower modules to focus on local details, while deeper modules integrate global behavior, thereby effectively capturing crawling phenomena at different time scales.

[0023] The first convolutional module uses N learnable convolutional kernels, each generating an independent feature map. The feature map is the output obtained after the input velocity sequence is processed by the convolutional kernel in the convolutional neural network. This layer primarily captures short-term, local velocity abrupt changes. The batch normalization layer performs channel-by-channel normalization on the N feature maps output from the previous layer. Specifically, for each feature map, its mean and variance on a mini-batch are calculated, and all values ​​are adjusted to a distribution with a mean close to 0 and a variance close to 1, accelerating model training convergence. The linear rectified activation layer performs a non-linear transformation on each value after batch normalization. For each value in the input, if it is negative, it is changed to 0; if it is positive or zero, it remains unchanged, mitigating the gradient vanishing problem and enhancing the network's expressive power. The max pooling layer sets the pooling window size and stride. This layer takes the maximum value between every two adjacent time points as the output, halving the time length of the feature map and reducing the computational cost of subsequent layers. The output of the first convolutional module is N feature maps.

[0024] The second convolutional module receives multiple feature maps output from the pooling layer of the first module as input. The convolutional layer uses 2N convolutional kernels, each generating an independent feature map. The padding method is set to be the same, i.e., zeros are padded at both ends of the sequence to ensure the output feature map's duration is the same as the input. Combined with the previous pooling layer which halves the feature map's duration, the effective receptive field gradually expands, covering a longer time range. This allows for the combination of the initial features extracted by the first module to form short waveform segments, i.e., brief pauses caused by rapid speed changes in the crawling phenomenon. The batch normalization layer performs channel-by-channel normalization on the 2N output feature maps. For each feature map, its mean and variance on a mini-batch are calculated, and the values ​​are adjusted to a distribution with a mean close to 0 and a variance close to 1, accelerating model training convergence. Next, the linear rectified activation layer performs a non-linear transformation on each batch-normalized value. For each input value, if it is negative, it is changed to 0; if it is positive or zero, it alleviates the vanishing gradient problem and enhances the network's expressive power. A max-pooling layer is set up with a pooling window size and stride of 0. The maximum value between any two adjacent time points is taken as the output, thereby halving the time length of the feature map and reducing the computational cost of subsequent layers. The output of the second convolutional module is 2N feature maps.

[0025] The third convolutional module receives 2N feature maps as input from the pooling output of the second module. The convolutional layer uses 4N convolutional kernels, each generating an independent feature map. The padding method is set to be the same, i.e., zeros are padded at both ends of the sequence to ensure the output feature map has the same duration as the input. After the first two pooling layers, its effective receptive field covers the velocity sequence further. Therefore, this layer can combine the local velocity abrupt changes and short waveform fragments extracted by the first two layers into a complete representation of the crawling event. Subsequently, the batch normalization layer performs channel-by-channel normalization on the output feature maps: for each feature map, its mean and variance on a mini-batch are calculated, and the values ​​are adjusted to a distribution with a mean close to 0 and a variance close to 1, which accelerates model training convergence. The linear rectified activation layer performs a non-linear transformation on each value after batch normalization. For each value in the input, if it is negative, it is changed to 0; if it is positive or zero, it alleviates the gradient vanishing problem and enhances the network's expressive power. A max-pooling layer is set up with a pooling window size and stride of 0. The maximum value between any two adjacent time points is used as the output, thus halving the time length of the feature maps and reducing the computational cost of subsequent layers. The output of the third convolutional module is 4N feature maps. These feature maps are then fed into a global average pooling layer and a fully connected classification layer to determine the probability of crawling events.

[0026] The global average pooling layer calculates the average of all values ​​in each feature map along the time axis from the feature map output by the last convolutional module, thus compressing each feature map into a scalar. All scalars are combined into a fixed-length feature vector, which is input to the first fully connected layer, linearly mapping it to a multidimensional space to further integrate global features. A linear rectified activation function is then applied, setting negative numbers to zero and leaving positive numbers unchanged, introducing non-linear expressive power. A dropout layer is used, randomly setting the output of some neurons to zero with a 50% probability during training to prevent the network from becoming dependent on specific neurons, thereby enhancing generalization ability and reducing overfitting. Finally, a second fully connected layer maps the multidimensional features to a 2D space and performs exponential normalization using a flexible maximum activation function, outputting probability distributions for two categories: "crawl" and "normal," respectively.

[0027] The model training employs a supervised learning paradigm. The network minimizes the cross-entropy loss between the predicted probability distribution and the true label using the Adam optimizer. The Adam optimizer is an adaptive learning rate gradient descent algorithm that dynamically adjusts the learning rate based on the historical gradients of each parameter, enabling it to converge quickly and stably to a smaller value of the loss function. The batch size is set to 128, and the iterations run for 25 epochs. After each epoch, the loss and accuracy are calculated on the validation set to monitor the model's generalization performance. After training, the model infers from real-time captured and preprocessed velocity segments, determining whether a crawling event has occurred based on the magnitude of the output probability value.

[0028] Step 3.3: During the monitoring phase, the system performs differentiation and moving average filtering on the measured raw displacement data to obtain the velocity sequence, which is then fed into the CNN model trained in Step 3.2 for judgment. The system outputs the crawling probability p for each data segment and sets a confidence threshold θ. When p ≥ θ, a crawling event is determined to have occurred within that time period. The system synchronously outputs a velocity-time curve and highlights the data segments determined to be crawling in red on the curve, thus achieving the monitoring and visualization of the crawling phenomenon.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention specifies that the measurement object is the final part of the linear shaft of a precision machine tool, which truly reflects the crawling characteristics. The high-precision measurement device is designed to ensure stable measurement under the actual processing conditions of the precision machine tool. The combination of traditional moving average filtering signal processing and deep learning algorithm results in high accuracy in identifying crawling phenomena. Attached Figure Description

[0030] Figure 1 Simplified structural diagram of a precision machine tool; Figure 2 Schematic diagram of the crawling phenomenon measurement system installation; Figure 3 8-Figure Simplified Diagram of Sensor Fastening Fixture Structure Figure 4 The raw data was directly used to plot the velocity. Figure 5 Velocity plotting is performed after moving average filtering; Figure 6 Signal processing and crawling recognition flowchart; Figure 7 Crawling phenomenon discrimination result diagram In the diagram: 1. Column; 2. Slide; 3. Slider; 4. Linear guide rail; 5. Mounting plate; 6. Precision adjustment platform; 7. Connecting plate; 8. Sensor fastening fixture; 9. Protective baffle; 10. Detection block. Detailed Implementation

[0031] The present invention will be further described below with reference to specific implementation examples.

[0032] This patent uses a certain type of precision machine tool as an example to describe a precision machine tool linear axis feed motion crawling measurement device and intelligent recognition method. The linear axis feed speed of the precision machine tool is set to 0.4 micrometers / second during the measurement process. To more clearly illustrate the invention, a detailed analysis of the invention will be provided below with reference to the accompanying drawings and embodiments.

[0033] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A high-precision measuring device for the linear axis feed motion crawling of a precision machine tool is disclosed. The device includes a mounting plate 5, a precision adjustment platform 6, a connecting plate 7, a sensor fastening fixture 8, an inner clamping arm 8-1, an outer clamping arm 8-2, a protective baffle 9, a detection block 10, and a sensor mounted on the sensor fastening fixture 8. To ensure sufficient stability during measurement, the device is connected to a column 1 via the mounting plate 5, with the remaining components mounted on the mounting plate 5. Specifically: Mounting plate 5 has a rectangular plate structure, which is used to vertically mount on column 1. All other components are mounted on mounting plate 5 to ensure stability during the measurement process.

[0034] Considering the small detection range of the selected capacitive displacement sensor (the sensor can only collect data when the distance between the sensor probe and the detected surface is less than 200 micrometers), to accurately set the initial gap between the sensor probe and the detected object to be less than 200 micrometers, the sensor fastening clamp 8 and the fastened sensor need to be able to make minor adjustments in the vertical direction. Therefore, a precision adjustment platform 6 is used for minor adjustments in the vertical direction. This precision adjustment platform 6 is equipped with an adjustment knob, and rotating the knob converts the rotational motion into linear motion through a high-precision threaded pair. The precision adjustment platform 6 is connected to the mounting plate 5. A connecting plate 7 is designed, with an overall rectangular plate structure. The right side of the connecting plate 7 is connected to the precision adjustment platform 6, and the left side is designed with two vertical bosses arranged in parallel. The distance between the two bosses is the width of the sensor fastening clamp 8, serving as a guide groove. The sensor fastening clamp 8 is placed in the guide groove and connected to the connecting plate 7 to ensure that the sensor fastening clamp 8 is placed vertically, thereby ensuring that the sensor is placed vertically. Rotating the adjustment knob drives the entire connecting plate 7 and the sensor fastening clamp 8, along with the fastened sensor, to make precise vertical displacement, thereby achieving precise adjustment of the initial gap.

[0035] A sensor clamping fixture 8 is designed to secure the sensor. Its structure consists of two arms, an inner arm 8-1 and an outer arm 8-2. The inner arm 8-1 is longer than the outer arm 8-2. The inner arm 8-1 connects to the guide groove of the connecting plate 7. Each arm has a semi-cylindrical groove perpendicularly designed on its inner side to hold the cylindrical sensor probe. An optical hole is designed on the outer arm 8-2, and a threaded hole is designed at the corresponding position on the inner arm 8-1. A bolt passes through the optical hole on the outer arm 8-2 and is screwed into the threaded hole on the inner arm 8-1. During bolt tightening, the two arms gradually close together, generating clamping force to fix the sensor probe. A threaded hole is designed at the top of the sensor clamping fixture 8, which cooperates with the protective baffle 9 for fine adjustments.

[0036] To prevent sudden changes in the linear axis feed speed of a precision machine tool from causing the horizontal portion of the detection block 10 to collide with the sensor probe, a protective baffle 9 is designed to prevent the sensor probe from being struck first during a sudden change in linear axis feed speed. Specifically: The protective baffle 9 is a rectangular plate structure, placed horizontally above and connected to the connecting plate 7. A through-hole, slightly larger than the sensor probe diameter, is designed in the center of the protective baffle 9, allowing the sensor probe to pass through and measure the horizontal portion of the detection block 10. During sensor installation, manual adjustment ensures the sensor probe is slightly below the upper surface of the protective baffle 9. Therefore, in the event of a sudden change in the linear axis feed speed causing an impact, the horizontal portion of the detection block 10 above the protective baffle 9 first impacts the upper surface of the protective baffle 9, preventing direct impact on the sensor probe below, thus effectively protecting the sensor probe.

[0037] Furthermore, to accurately set and maintain the aforementioned protective positional relationship, a fine-tuning mechanism is provided on the protective baffle 9. The right end of the protective baffle 9 has two holes, one on the left and one on the right. The left hole is a through-hole, which engages with the threaded hole above the sensor fastening fixture 8 for minor adjustments. A bolt is screwed through the through-hole into the threaded hole above the sensor fastening fixture 8. During tightening, the main body of the sensor fastening fixture 8 and the fastened sensor probe are pulled upwards relative to the protective baffle 9, bringing the sensor probe closer to the upper surface of the protective baffle 9. The right hole is a through-hole, which is screwed into the threaded hole above the protective baffle 9. The bolt is screwed in to a depth exceeding the depth of the threaded hole, with the end of the bolt pressing against the upper surface of the sensor fastening fixture 8. During tightening, the main body of the sensor fastening fixture 8 and the fastened sensor probe are pressed downwards relative to the protective baffle 8, moving the sensor probe away from the upper surface of the protective baffle 9. By alternately adjusting the left and right bolts, ensure that the upper surface of the protective baffle 9 is always above the sensor probe to protect the sensor probe.

[0038] Analyzing the linear axis feed motion transmission chain of a precision machine tool, the transmission motion is determined by a motor rotating to drive a lead screw, which in turn drives a lead screw nut to move along the linear guide rail 4, thereby moving the slide 2 connected to the lead screw nut and slider 3. It is clear that the final component reflecting the crawling phenomenon is the slide 2. Therefore, the displacement of the slide 2 needs to be detected during the testing process. Considering the limitations of the testing space and the requirement for sufficient flatness of the tested surface, a detection block 10 is designed as the actual testing object. This L-shaped plate is divided into a vertical section and a horizontal section. The vertical section is connected to the side of the slide 2 and is located on both sides of the connecting plate 7 with the sensor fastening fixture 8, maintaining a certain distance from the connecting plate 7. The horizontal section is located above the protective baffle 9. When the linear axis moves vertically downwards, the horizontal section of the detection block 10 gradually approaches the protective baffle 9. When the distance between the detection block 10 and the sensor probe fastened by the sensor fastening fixture 8 below the upper surface of the protective baffle 9 is less than the sensor's detection range, the sensor detects the displacement of the horizontal section of the detection block 10.

[0039] A method for intelligent identification of crawling measurement of linear axis feed motion in precision machine tools, based on the aforementioned measuring device, includes the following steps: The first step is to install the measuring device. Once the measuring device is fully installed, measurements can begin. Specifically: Step 1.1: Use bolts to connect the mounting plate 5 to the column 1 to obtain the fixed mounting plate 5; Step 1.2: For the fixed mounting plate 5, use bolts to connect the precision adjustment platform 6 to the mounting plate 5 to obtain the fixed precision adjustment platform 6; Step 1.3: For the fixed precision adjustment platform 6, use bolts to connect the right side of the connecting plate 7 to the precision adjustment platform 6 to obtain the fixed connecting plate 7.

[0040] Step 1.4: For the fixed connecting plate 7, use bolts to initially fix the sensor fastening clamp 8 to the left guide groove of the connecting plate 7, but do not lock it, so as to facilitate subsequent adjustments, and obtain the initially fixed sensor fastening clamp 8.

[0041] Step 1.5: For the fixed connecting plate 7, use bolts to connect the protective baffle 9 to the upper end of the connecting plate 7 to obtain the fixed protective baffle 9.

[0042] Step 1.6: Use bolts to pass through the light hole of the outer clamping arm 8-2 and screw them into the threaded hole of the inner clamping arm 8-1 to gradually bring the two clamping arms together to generate clamping force to fix the sensor probe.

[0043] Step 1.7: Use bolts to alternately adjust the light hole arranged on the upper right end of the protective baffle 9 and the threaded hole designed on the upper part of the sensor fastening clamp 8, as well as the threaded hole on the right side of the light hole, so that the sensor probe is located below the upper surface of the protective baffle 9. Then, finally fix the sensor fastening clamp 8, which was initially fixed in step 1.4, to the connecting plate 7.

[0044] Step 1.8: Use bolts to connect the vertical part of the detection block 10 to the side of the slide 2, and the horizontal part is located above the protective baffle 9, finally obtaining the measuring device after the whole assembly is completed.

[0045] The second step involves adjusting the parameters of the precision machine tool control panel after the measuring device has been fully installed in the first step to obtain measurement data; specifically: Step 2.1: Change the linear axis feed motion parameters on the precision machine tool control panel to make the linear axis feed downwards. The slide 2 on the linear axis also feeds downwards, thereby driving the detection block 10 connected to the side of the slide 2 to feed downwards.

[0046] Furthermore, since a capacitive displacement sensor with a range of 200 micrometers is selected, the sensor can collect data when the distance between the sensor probe and the measured surface is less than 200 micrometers. When the distance between the horizontal part of the visual detection block 10 and the sensor probe is about 5mm but the sensor does not show a reading, the linear axis movement is stopped, and the adjustment knob of the precision adjustment platform 6 is adjusted. When the adjustment is made so that the sensor shows a reading, the adjustment knob is stopped.

[0047] Step 2.2: When the sensor displays a reading, readjust the machine tool control panel to change the feed speed to 0.4 micrometers / second, the feed amount to 12 micrometers, set the motion to reciprocating motion, start the linear axis to start feeding, and control the sensor to start collecting data, finally obtaining the measurement data.

[0048] The third step, based on the measurement data obtained in the second step, involves preliminary data processing and training a deep learning algorithm to accurately identify crawling phenomena. The identification process is as follows: Figure 5 The signal processing and crawling recognition flowchart is shown below. Specifically: Step 3.1: Differentiate the original displacement data obtained in step 2 to obtain the original velocity data. Then, apply a moving average filter of N=100 to the velocity data to obtain noise-removed velocity data. The filtered signal at the current moment... It is given by the following formula: in, It is the first one The filtered output value The length of the sliding window. The first input signal Each sample value. This processing can significantly improve the signal-to-noise ratio, providing stable and smooth input data for subsequent feature recognition.

[0049] Depend on Figure 3 Directly plotting velocity from raw data and Figure 4 The velocity plot after applying the moving average filter shows that the moving average filter significantly reduces the impact of noise.

[0050] Step 3.2: Construct a temporal classification model based on a Convolutional Neural Network (CNN). This model takes the velocity sequence filtered by the moving average in Step 3.1 as input. The velocity sequence is divided into several velocity segments with a fixed window length (256 columns), each segment being an independent sample. To generate a label for each sample, the velocity difference between the same velocity segment under crawling and non-crawling conditions is calculated. This velocity difference is compared to a set difference threshold (0.1 μm / s). If the difference is greater than the threshold, crawling is considered to exist; if it is less than the threshold, non-crawling is considered to exist. The true state ("crawling" or "normal") at each sampling time is determined and recorded as 0 or 1. Then, the label for each velocity segment is determined by the state of the majority of sampling points within that segment. If the majority of points are crawling, the label is 1 (crawling); otherwise, it is 0 (normal). All labeled samples are randomly divided into a training set and a validation set in an 8:2 ratio, with the validation set used to monitor the model's generalization performance. Design a one-dimensional convolutional neural network for temporal classification, consisting of three convolutional modules, a global average pooling layer, and a fully connected classification layer connected in series. Each convolutional module includes a convolutional layer, a batch normalization layer, a linear rectified activation layer, and a max pooling layer. Specifically: In velocity sequences, crawling is characterized by sudden velocity changes within a short time window, with complete crawling events and their intervals typically lasting from tens of milliseconds to several seconds. Convolutional layers are the core component of convolutional modules, their function being to extract local features of the velocity sequence using learnable convolutional kernels. As the specific feature extraction operator within a convolutional layer, a single convolutional layer, having only one set of learnable kernels, is limited by its fixed receptive field (i.e., the time span covered by the kernel on the input velocity sequence), and can only extract limited details within that receptive field, making it difficult to simultaneously capture both local transient features and global event trends. Therefore, this network employs three convolutional modules, progressively increasing the receptive field layer by layer. This allows shallower modules to focus on local details, while deeper modules integrate global behavior, thereby effectively capturing crawling phenomena at different time scales.

[0051] The first convolutional module uses 16 learnable convolutional kernels with a kernel size of [1, 31]. Each kernel generates an independent feature map, resulting in 16 feature maps. The feature map is the output of the input velocity sequence after processing by the convolutional kernels in the convolutional neural network. The padding method is set to be the same, i.e., zeros are padded at both ends of the sequence to ensure that the time length of the output feature map is the same as the input (still 256 columns). This layer mainly captures short-term, local velocity mutations. The batch normalization layer performs channel-by-channel normalization on the 16 feature maps output from the previous layer. Specifically, for each feature map, its mean and variance on a mini-batch are calculated, and then all values ​​are adjusted to a distribution with a mean close to 0 and a variance close to 1, which can accelerate model training convergence. The linear rectified activation layer performs a non-linear transformation on each value after batch normalization. For each value in the input, if it is negative, it is changed to 0; if it is positive or zero, it remains unchanged, which can alleviate the gradient vanishing problem and enhance the network's expressive power. The max-pooling layer has a pooling window size of [1,2] and a stride of [1,2]. This layer takes the maximum value between any two adjacent time points as its output, thereby halving the length of the feature map (from 256 columns to 128 columns) and reducing the computational cost of subsequent layers. The output of the first convolutional module is 16 feature maps, each with 128 columns.

[0052] The second convolutional module receives 16 feature maps (each with 128 columns) as input from the pooling output of the first module. The convolutional layer uses 32 kernels of size [1, 15]. Each kernel generates an independent feature map, resulting in 32 output feature maps. The padding method is the same: zeros are padded at both ends of the sequence to ensure the output feature map's duration is the same as the input (still 128 columns). Compared to the [1, 31] kernels of the first module, this module uses smaller [1, 15] kernels. Combined with the halved duration of the feature maps resulting from the previous pooling layer, the effective receptive field gradually expands, covering a longer time range. This allows the module to combine the initial features extracted by the first module to form short-time waveform segments, representing the brief pauses caused by rapid speed changes during crawling. The batch normalization layer performs channel-wise normalization on the 32 output feature maps. For each feature map, it calculates its mean and variance on the mini-batch samples and adjusts the values ​​to a distribution with a mean close to 0 and a variance close to 1, which can accelerate model training convergence. Next, the linear rectified activation layer performs a non-linear transformation on each value after batch normalization. For each value in the input, if it is negative, it is changed to 0; if it is positive or zero, it can alleviate the gradient vanishing problem and enhance the network's expressive power. A max pooling layer is set with a pooling window size of [1, 2] and a stride of [1, 2]. The maximum value between every two adjacent time points is taken as the output, thereby halving the time length of the feature map (from 128 columns to 64 columns) and reducing the computational cost of subsequent layers. The output of the second convolutional module is 32 feature maps, each with 64 columns.

[0053] The third convolutional module receives 32 feature maps (each with 64 columns) as input from the pooling output of the second module. The convolutional layer uses 64 kernels of size [1, 7]. Each kernel generates an independent feature map, resulting in a total of 64 output feature maps. The padding method is the same: zeros are padded at both ends of the sequence to ensure the output feature map has the same duration as the input (still 64 columns). Compared to the [1, 15] kernels in the second module, this module uses smaller [1, 7] kernels, and after the first two pooling layers, its effective receptive field covers the velocity sequence further. Therefore, this layer can combine the local velocity abrupt changes and short waveform fragments extracted by the first two layers into a complete representation of the crawling event. Subsequently, the batch normalization layer performs channel-by-channel normalization on the 64 output feature maps: for each feature map, its mean and variance on the mini-batch samples are calculated, and the values ​​are adjusted to a distribution with a mean close to 0 and a variance close to 1, which can accelerate model training convergence. The linear rectified activation layer performs a non-linear transformation on each value after batch normalization. For each value in the input, if it is negative, it is changed to 0; if it is positive or zero, it can alleviate the gradient vanishing problem and enhance the network's expressive power. A max pooling layer is set with a pooling window size of [1, 2] and a stride of [1, 2]. The maximum value between every two adjacent time points is taken as the output, thereby halving the time length of the feature map (from 64 columns to 32 columns) and reducing the computational cost of subsequent layers. The output of the third convolutional module is 64 feature maps, each with 32 columns. These feature maps will then be fed into a global average pooling layer and a fully connected classification layer to complete the probability discrimination of crawling events.

[0054] The global average pooling layer calculates the average of all values ​​in each feature map along the time axis from the feature map output by the last convolutional module, thus compressing each feature map into a scalar. All scalars are combined into a fixed-length 64-dimensional feature vector, which is input to the first fully connected layer, linearly mapping it to a 64-dimensional space to further integrate global features. A linear rectified activation function is then applied, setting negative numbers to zero and leaving positive numbers unchanged, introducing non-linear expressive power. A dropout layer is used, randomly setting the output of some neurons to zero with a 50% probability during training to prevent the network from becoming dependent on specific neurons, thereby enhancing generalization ability and reducing overfitting. Finally, a second fully connected layer maps the 64-dimensional features to a 2-dimensional space and performs exponential normalization using a flexible maximum activation function, outputting probability distributions for two categories: "crawl" and "normal," respectively.

[0055] The model training employs a supervised learning paradigm. The network minimizes the cross-entropy loss between the predicted probability distribution and the true label using the Adam optimizer. The Adam optimizer is an adaptive learning rate gradient descent algorithm that dynamically adjusts the learning rate based on the historical gradients of each parameter, enabling it to converge quickly and stably to a smaller value of the loss function. The batch size is set to 128, and the iterations run for 25 epochs. After each epoch, the loss and accuracy are calculated on the validation set to monitor the model's generalization performance. After training, the model infers from real-time captured and preprocessed velocity segments, determining whether a crawling event has occurred based on the magnitude of the output probability value.

[0056] Step 3.3, the online monitoring phase, involves the system segmenting the real-time collected velocity sequences into segments of equal window length, normalizing them, and then feeding them into the CNN model trained in Step 3.2 for judgment. The system outputs the crawling probability p for each data segment, setting a confidence threshold θ = 0.9 for classifying crawling as a crawling phenomenon. When p ≥ θ, a crawling event is determined to have occurred within that time period. The system simultaneously outputs a velocity-time curve and highlights the data segments identified as crawling in red on the curve, achieving monitoring and visualization of the crawling phenomenon.

[0057] like Figure 6 After training, the system performs discrimination and visualizes the output. This primarily includes: the velocity-time curve obtained through differentiation and moving average filtering, and the state discrimination results output by the CNN convolutional neural network model. The system marks the time periods that the model classifies as "crawling" in red on the velocity curve, thus intuitively displaying the occurrence time, duration, and corresponding velocity fluctuation characteristics of crawling events. This achieves a complete monitoring process from data acquisition and deep learning analysis to state visualization.

[0058] The above embodiments are merely illustrative of the implementation methods of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A precision machine tool linear axis feed motion crawling measurement device, characterized in that, The measuring device includes a mounting plate (5), a precision adjustment platform (6), a connecting plate (7), a sensor fastening fixture (8), a protective baffle (9), a detection block (10), and a sensor mounted on the sensor fastening fixture (8). It is connected to the column (1) of the precision machine tool via the mounting plate (5), and the remaining parts of the device are mounted on the mounting plate (5). Specifically: The precision adjustment platform (6) enables minute adjustments in the vertical direction of the sensor fastening fixture (8) and the fastened sensor; the precision adjustment platform (6) is connected to the mounting plate (5) and is capable of minute adjustments in the vertical direction. The connecting plate (7) is a rectangular plate structure. The right side of the connecting plate (7) is connected to the precision adjustment platform (6). Two vertical bosses are designed on the left side. The two bosses are arranged in parallel and the space between the two bosses serves as a guide groove for installing the sensor fastening fixture (8). The sensor is vertically installed on the sensor fastening fixture (8). Through the precision adjustment platform (6), the connecting plate (7) and the sensor can be moved vertically to achieve precise adjustment of the initial gap. The protective baffle (9) is a rectangular plate structure, placed horizontally, and located above the two protrusions of the connecting plate (7). It is used to prevent the horizontal part of the detection block (10) from hitting the sensor probe due to sudden changes in the feed speed of the linear axis of the precision machine tool. The protective baffle (9) has a through-hole in the middle, with a diameter larger than that of the sensor probe. The sensor probe can pass through the hole to measure the horizontal part of the detection block (10). During installation, the sensor probe is lower than the upper surface of the protective baffle (9). The detection block (10) is the actual object to be detected. It is an L-shaped plate divided into a vertical part and a horizontal part. The vertical part is connected to the side of the slide (2) and the sensor fastening clamp (8) is located on both sides of the connecting plate (7). The vertical part does not contact the connecting plate (7), and the horizontal part is located above the protective baffle (9). When the linear shaft moves vertically downward, the horizontal part of the detection block (10) connected to the side of the slide (2) gradually approaches the protective baffle (9). When the distance between the detection block (10) and the sensor probe is less than the sensor detection range, the sensor performs displacement detection on the horizontal part of the detection block (10) and thus obtains the displacement of the slide (2).

2. The precision machine tool linear axis feed motion crawling measuring device according to claim 1, characterized in that, The sensor is finely adjusted to achieve an initial gap of less than 200 micrometers between the sensor probe and the object being detected.

3. The precision machine tool linear axis feed motion crawling measurement device according to claim 2, characterized in that, The mounting plate (5) is a rectangular plate structure, which is vertically installed on the column (1), and the other parts are installed on the mounting plate (5).

4. The precision machine tool linear axis feed motion crawling measuring device according to claim 3, characterized in that, The precision adjustment platform (6) is equipped with an adjustment knob. Rotating the knob converts the rotational motion into linear motion through a high-precision threaded pair. Rotating the adjustment knob drives the entire connecting plate (7), sensor fastening fixture (8), and sensor to make precise vertical displacement, thereby achieving precise adjustment of the initial gap.

5. A precision machine tool linear axis feed motion crawling measuring device according to claim 4, characterized in that, The specific structure of the sensor fastening clamp (8) is as follows: It consists of two inner and outer arms. The inner arm 8-1 is longer than the outer arm 8-2. The inner arm 8-1 is connected to the guide groove of the connecting plate (7). A semi-cylindrical groove is designed vertically on the inner side of each of the two arms to place the cylindrical probe of the sensor. A light hole is designed on the outer arm 8-2, and a threaded hole is designed at the same position on the inner arm. The bolt passes through the light hole of the outer arm and is screwed into the threaded hole of the inner arm. During the tightening of the bolt, the two arms gradually move closer to fix the probe of the sensor. The sensor fastening clamp (8) has a threaded hole on top, which is used to cooperate with the protective baffle (9) for fine adjustment.

6. The precision machine tool linear axis feed motion crawling measuring device according to claim 5, characterized in that, The protective baffle (9) is provided with a fine-tuning mechanism, and the right end of the protective baffle (9) is designed with two holes, one on the left and one on the right. The left hole is a through-hole for fine adjustment. The left hole is used to make slight adjustments by engaging with the threaded hole on the upper part of the sensor fastening clamp (8). A bolt is screwed through the hole on the upper part of the protective baffle (9) and into the threaded hole on the upper part of the sensor fastening clamp (8). During the tightening process, the sensor fastening clamp (8) and the sensor probe being fastened are pulled upward relative to the protective baffle (9) so that the sensor probe is close to the upper surface of the protective baffle (9). The right-side hole is a through threaded hole. A bolt is screwed into the threaded hole above the protective baffle (9). The depth of the bolt exceeds the depth of the threaded hole. The end of the bolt rests on the upper surface of the sensor fastening clamp (8). During the tightening process, the sensor fastening clamp (8) and the fastened sensor probe are pressed downward relative to the protective baffle (8), so that the sensor probe is away from the upper surface of the protective baffle (9). By alternately adjusting the left and right bolts, the upper surface of the protective baffle (9) is ensured to be above the sensor probe to protect the sensor probe.

7. A method for intelligent recognition of crawling measurement of linear axis feed motion in precision machine tools, characterized in that, Based on the measuring device according to any one of claims 1-6, the method includes the following steps: The first step is to install the measuring device. Once the measuring device is fully installed, measurements can begin. Specifically: Step 1.1: Connect the mounting plate (5) to the column (1) to obtain the fixed mounting plate (5); Step 1.2: For the fixed mounting plate (5), connect the precision adjustment platform (6) to the mounting plate (5) to obtain the fixed precision adjustment platform (6); Step 1.3: For the fixed precision adjustment platform (6), connect the right side of the connecting plate (7) to the precision adjustment platform (6) to obtain the fixed connecting plate (7); Step 1.4: For the fixed connecting plate (7), the sensor fastening clamp (8) is initially fixed to the left guide groove of the connecting plate (7), but not locked, so as to facilitate subsequent adjustment, and the sensor fastening clamp (8) is initially fixed. Step 1.5: For the fixed connecting plate (7), connect the protective baffle (9) to the upper end of the connecting plate (7) to obtain the fixed protective baffle (9); Step 1.6: Use bolts to pass through the light hole of the outer arm and screw into the threaded hole of the inner arm to fasten the sensor clamp (8), so that the two arms gradually come together to generate clamping force to fix the sensor probe. Step 1.7: Use bolts to alternately adjust the light hole arranged on the upper right end of the protective baffle (9) and the threaded hole designed on the upper part of the sensor fastening clamp (8), as well as the threaded hole on the right side of the light hole, so that the sensor probe is located below the upper surface of the protective baffle (9). Then, finally fix the sensor fastening clamp (8) that was initially fixed in Step 1.4 to the connecting plate (7). Step 1.8: Connect the vertical part of the detection block (10) to the side of the slide (2), and the horizontal part is located above the protective baffle (9) to obtain the measuring device after the whole assembly is completed; The second step involves adjusting the machine tool control panel parameters based on the completed measuring device from the first step to obtain measurement data; specifically: Step 2.1: Change the motion parameters of the linear axis in the machine tool control panel to make the linear axis feed downwards. The slide (2) on the linear axis also feeds downwards, driving the detection block (10) connected to the side of the slide (2) to feed downwards. Step 2.2: When the sensor displays a reading, readjust the machine tool control panel to the parameters to be measured, start the linear axis to begin feeding, and control the sensor to start collecting data, finally obtaining the raw displacement data; The third step, based on the raw displacement data obtained in the second step, involves preliminary data processing and training a deep learning algorithm to accurately identify the crawling phenomenon. Specifically: Step 3.1: Differentiate the original displacement data obtained in the second step to obtain the original velocity data, and then perform moving average filtering on the velocity data to obtain the velocity sequence with noise removed. The filtered velocity sequence at the current moment is then obtained. in, It is the first one Filtered velocity sequence The length of the sliding window. The first input signal Each sample value; Step 3.2: Construct a temporal classification model based on a convolutional neural network (CNN). This temporal classification model takes the velocity sequence after the moving average filtering in Step 3.1 as input and compares it with the velocity when the straight axis is running normally, i.e., without crawling. The model automatically extracts local temporal features through convolutional layers and outputs the probability of crawling events after integration by fully connected layers. Step 3.3, the online monitoring phase: the system segments the real-time collected velocity sequence into segments with the same window length, normalizes them, and then feeds them into the time series classification model trained in Step 3.2 for judgment, outputting the crawling probability p for each segment of data, setting a confidence threshold θ, and determining that a crawling event has occurred within that time period when p ≥ θ; the system synchronously outputs the velocity-time curve, and highlights the data intervals determined to be crawling in red on the curve, realizing the monitoring and visualization of the crawling phenomenon.

8. The intelligent identification method for measuring the crawling motion of a linear axis in a precision machine tool according to claim 7, characterized in that, In step 2.1, the base is a capacitive displacement sensor with a range of 200 micrometers. When the distance between the sensor probe and the measured surface is less than 200 micrometers, the sensor can collect data. When the horizontal part of the visual inspection block (10) is close to the sensor probe but the sensor does not show a reading, the linear axis movement is stopped, and the adjustment knob of the precision adjustment platform (6) is adjusted. When the adjustment is made so that the sensor shows a reading, the adjustment knob is stopped.

9. The intelligent identification method for measuring the crawling motion of a linear axis in a precision machine tool according to claim 8, characterized in that, Step 3.2 specifically involves: The temporal classification model takes the velocity sequence after moving average filtering in step 3.1 as input; the velocity sequence is divided into several velocity segments according to a fixed window length, and each segment is an independent sample; to generate a label for each sample, the velocity difference of the same velocity segment under crawling conditions and non-crawling conditions is calculated, and the velocity difference is compared with a set difference threshold. If the difference is greater than the difference threshold, crawling is considered to exist and recorded as 0; if the difference is less than the difference threshold, non-crawling is considered to exist and normal and recorded as 1; all labeled samples are divided into training set and validation set; a one-dimensional convolutional neural network for temporal classification is designed, which consists of three convolutional modules, a global average pooling layer and a fully connected classification layer connected in series; each convolutional module contains a convolutional layer, a batch normalization layer, a linear rectified activation layer and a max pooling layer; The model training adopts a supervised learning paradigm; the network minimizes the cross-entropy loss between the predicted probability distribution and the true label through the Adam optimizer; After training, the model infers from real-time collected and preprocessed velocity segments and determines whether a crawling event has occurred based on the magnitude of the output probability value.

10. The intelligent identification method for measuring the crawling motion of a precision machine tool linear axis feed motion according to claim 9, characterized in that, The difference threshold is 0.1 micrometers / second; the confidence threshold θ is 0.9.

Images