Method and apparatus for detecting abnormal press fitting of power transmission disc assembly

Abstract

The application discloses a power transmission disc assembly press-fit abnormality detection method and device, the method comprises the following steps: acquiring a displacement sequence and a force value sequence of a press-fit effective section, and generating a force value increment sequence; inputting the three types of sequences into a shallow feature extraction network of an abnormality detection model to obtain a basic feature sequence; locating a sudden drop valley node in the force value increment sequence, determining an instantaneous stiffness sequence according to the force value sequence and the displacement sequence, and dividing a left time window and a right time window corresponding to the valley node according to the instantaneous stiffness sequence; extracting a bidirectional deformation index and an irreversible plastic work feature based on the two time windows, generating a topography comparison coefficient, and weighting the basic feature sequence; and performing global time sequence relationship aggregation on the weighted feature sequence to output a classification determination result for controlling a press-fit device. The application can improve the distinguishing accuracy of press-fit abnormalities such as stick-slip friction and micro-cracks, and realize online detection and device control in the press-fit process.

Description

Technical Field

[0001] This application relates to the field of intelligent manufacturing and quality inspection technology for automotive parts, specifically to a method and equipment for detecting abnormal pressing of a power transmission disc assembly. Background Technology

[0002] With the development of the automotive industry, press fitting is widely used in the assembly of powertrain disc assemblies. This process is mainly applied in the manufacturing of clutch disc hub assemblies, interference fit assemblies between flexible transmission discs and flanges, and other similar components in automatic transmissions. In actual workshop processing environments, servo press fitting equipment is typically used to monitor press fitting quality. This equipment is equipped with piezoelectric force sensors and high-precision grating ruler displacement sensors to collect real-time data on the spindle axial press fitting force and pressing stroke.

[0003] However, existing methods for detecting abnormalities in press-fitting quality face significant technical bottlenecks in practical applications. When relative sliding occurs between the interference fit surfaces of metal stampings and machined parts, the differences in surface roughness and uneven distribution of the anti-rust oil film easily trigger the physical phenomenon of stick-slip friction. This phenomenon produces high-frequency force fluctuations that are remarkably similar to those experienced when a part fractures due to stress and microcracks. Relying on conventional data statistics or basic curve analysis often makes it difficult to effectively distinguish normal stick-slip friction fluctuations from actual abnormal microcrack propagation, easily leading to misjudgments of good products or missed detection of defective products.

[0004] Furthermore, different types of sensors in industrial settings operate on different measurement principles, and their electrical signal transmission and conversion circuits exhibit inherent delays at the microsecond level. Directly bundling data collected at the same timestamp can lead to microscopic distortion of the mechanical curves. Introducing external auxiliary detection equipment, such as acoustic or visual sensors, to improve defect identification accuracy inevitably increases hardware manufacturing costs. Therefore, there is an urgent need in this field for a method that can decouple complex waveforms and accurately identify pressing anomalies without requiring additional hardware. Summary of the Invention

[0005] In view of this, this disclosure provides a method and equipment for detecting abnormal pressing of power transmission disc assemblies, which can alleviate, reduce or even eliminate at least the above-mentioned problems.

[0006] According to one aspect of this application, a method for detecting press-fitting abnormalities in a power transmission disc assembly is provided, comprising: S1, obtain the displacement sequence and force value sequence of the effective section of the press fitting, and generate the corresponding force value increment sequence; S2, input the displacement sequence, force value sequence and force value increment sequence into the shallow feature extraction network of the anomaly detection model to extract features and obtain the basic feature sequence; S3, obtaining the weighted feature sequence by performing operations through the attention extraction layer of the anomaly detection model, specifically including: Locate the trough nodes in the force increment sequence that exhibit a sudden drop. The instantaneous stiffness sequence is determined based on the force value sequence and the displacement sequence, and the left and right time windows corresponding to the trough nodes are defined based on the instantaneous stiffness sequence. Based on the left and right time windows, a bidirectional deformation index is extracted, and an irreversible plastic work characteristic reflecting the energy dissipation state is extracted. Based on the deformation index and the irreversible plastic work feature, a morphology comparison coefficient is generated, and the morphology comparison coefficient is used to perform node weighting on the basic feature sequence; S4. The anomaly detection model is used to perform global temporal relationship aggregation on the weighted feature sequence of nodes, and output the classification judgment result for controlling the pressing equipment.

[0007] Optionally, obtaining the displacement sequence includes: extracting abrupt change points in the response slopes of the two types of sensors used for data acquisition during the initial stage of macroscopic contact; calculating a time difference parameter based on the abrupt change points; performing time axis compensation on the initially acquired displacement data points using interpolation based on the time difference parameter; and outputting the aligned displacement sequence.

[0008] Optionally, the left and right time windows corresponding to the trough nodes are defined based on the instantaneous stiffness sequence. Specifically, this includes: tracing back the numerical state of the instantaneous stiffness sequence to the left along the time axis, defining the time node corresponding to the first stable value exceeding the set structural stiffness threshold as the starting point of the left time window; searching the numerical state of the instantaneous stiffness sequence to the right along the time axis, defining the node where the value recovers to the range of the structural stiffness threshold and the single-step displacement increment exceeds the minimum displacement setting value as the ending point of the right time window; when the value fails to recover to the range of the structural stiffness threshold within the set maximum search span, the cutoff position of the maximum search span is forcibly defined as the ending point.

[0009] Optionally, the bidirectional deformation index includes a left drop deformation energy index and a right recovery deformation energy index; extracting the left drop deformation energy index specifically includes: for the sampling interval within the left time window, obtaining the absolute ratio between the force value change and the corresponding displacement change of adjacent samples, multiplying the absolute ratio by the velocity compensation coefficient, and then performing interval traversal accumulation to obtain the left drop deformation energy index; For the sampling interval within the right-hand time window, select the effective sampling interval that produces positive restoring force value, calculate the basic ratio of the force value change to the corresponding displacement change in the effective sampling interval, multiply the basic ratio by the velocity compensation coefficient, and then perform interval traversal and accumulation to obtain the right-hand restoring deformation energy index.

[0010] Optionally, extracting the right-side restoring deformation energy index specifically includes: for the sampling interval within the right-side time window, selecting the effective sampling interval that generates positive restoring force values, calculating the basic ratio of the force value change to the corresponding displacement change in the effective sampling interval, multiplying the basic ratio by the velocity compensation coefficient, and then performing interval traversal and accumulation to obtain the right-side restoring deformation energy index.

[0011] Optionally, extracting the speed compensation coefficient specifically includes: calculating the instantaneous pressing displacement rate value of the current corresponding stage; selecting the reference pressing speed value under the calibration condition; and using the ratio of the reference pressing speed value to the instantaneous pressing displacement rate value as the speed compensation coefficient of the corresponding sampling interval.

[0012] Optionally, the irreversible plastic work feature reflecting the energy dissipation state is extracted, specifically including: extracting discrete sampling node pairs within the closed-loop calculation interval formed by the starting point and the ending point; obtaining the average force distribution between adjacent sampling node pairs, and simultaneously obtaining the corresponding displacement increment range; multiplying and combining the average force distribution with the corresponding displacement increment range, performing discrete integral accumulation on the combined area parameters, and using the accumulated result representing the hysteresis loop dissipation feature as the irreversible plastic work feature.

[0013] Optionally, generating morphology comparison coefficients based on the deformation index and the irreversible plastic work feature specifically includes: extracting the basic ratio of the right-side recovery deformation energy index to the left-side drop deformation energy index; establishing a relative proportion state between the irreversible plastic work feature and a given benchmark plastic work constant; performing an exponential decay function transformation calculation combined with a negative decay factor using the relative proportion state as the core independent variable to obtain the exponential function value; and constructing the dimensionless morphology comparison coefficients by combining the basic ratio and the exponential function value through a multiplicative combination relationship.

[0014] Optionally, the anomaly detection model performs parameter training and optimization based on an error function that incorporates physical consistency constraints. The corresponding error function construction logic includes: using the distribution probability of the input sample in the target classification state category and combining it with cross-entropy to calculate the classification difference value of the benchmark; extracting the final force value parameter fed back by the input sample at the end of the pressing operation; when it is determined that the final force value parameter is lower than the set minimum qualified preload value lower limit, generating a physical penalty term based on the relative difference ratio of the final force value parameter deviating from the minimum qualified preload value lower limit; fusing the physical penalty term with the predicted probability value when the classification output is the normal state category to form a physical consistency deviation value; and weighting and summing the classification difference value and the physical consistency deviation value according to the weight allocation to serve as the total error basis for updating the network parameters of the anomaly detection model.

[0015] An electronic device is also provided, comprising a memory and a processor. The processor; the memory for storing a computer program executable by the processor; wherein the processor is configured to execute the computer program in the memory to implement the methods described herein.

[0016] According to the method of this application, by jointly analyzing the displacement sequence, force sequence and force increment sequence of the effective section of the press-fitting process, an attention extraction mechanism based on the mechanical characteristics of the press-fitting process is introduced into the anomaly detection model. The effective range of the force drop event is adaptively determined by using the instantaneous stiffness sequence. The morphology comparison coefficient is generated by combining the bidirectional deformation index after velocity compensation and the irreversible plastic work characteristics. Based on this, the local features are targeted and weighted, so that the model can distinguish between normal stick-slip friction and press-fitting anomalies such as microcracks and dimensional deviations from aspects such as stiffness recovery ability, deformation process and energy dissipation state, thereby reducing misjudgments caused by relying solely on the force fluctuation pattern.

[0017] Meanwhile, by compensating for the time axis of force and displacement signals, the impact of response delays of different sensors on feature extraction results is reduced; by global temporal relationship aggregation and physical consistency constraint training, the classification results take into account local mutation features, complete pressing trend and final preload requirements, thereby improving the accuracy, stability and cross-working condition adaptability of pressing anomaly detection. Attached Figure Description

[0018] This application will be more fully understood by referring to the following detailed description of specific embodiments in conjunction with the accompanying drawings, wherein the same reference numerals in the drawings refer to the same elements, wherein: Figure 1 A flowchart illustrating a method for detecting press-fitting anomalies in a power transmission disc assembly according to some embodiments of the present disclosure is shown.

[0019] Figure 2A schematic diagram illustrating the structure of an anomaly detection model according to some embodiments of the present disclosure is shown. Figure 3 A schematic diagram illustrating the execution flow of the attention extraction layer according to some embodiments of the present disclosure is shown.

[0020] Figure 4 A schematic diagram of the structure of an electronic device according to some embodiments of the present disclosure is shown. Detailed Implementation

[0021] Several embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to understand and implement the disclosure. However, the present disclosure can be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to make the disclosure comprehensive and complete, and to fully convey the scope of the disclosure to those skilled in the art. These embodiments are illustrative and not limiting. It should be noted that features from different embodiments can be combined arbitrarily without conflict.

[0022] The technical solution of this invention is mainly applied to the press-fit manufacturing of automotive powertrain disc assemblies. Automotive powertrain disc assemblies mainly include clutch disc hub assemblies in automatic transmissions, and interference fit assemblies between flexible transmission discs and flanges. In actual workshop processing environments, press-fitting equipment typically consists of a servo press body, a pressing spindle, an industrial control computer, and a programmable logic controller (PLC). To monitor press-fitting quality, the spindle is equipped with two types of core sensors: one is a piezoelectric force sensor for real-time acquisition of the axial pressing force of the spindle, and the other is a high-precision grating displacement sensor for synchronous acquisition of the spindle's pressing stroke. When the interference fit surfaces of the metal stamping part and the machined part slide relative to each other, due to differences in the micro-roughness of the component surfaces and the uneven distribution of the anti-rust oil film, stick-slip friction physical phenomena occur, generating high-frequency force fluctuations that are extremely similar to the fracture of micro-cracks in the parts. This solution relies on edge computing nodes or on-site industrial control computers as the execution subject, and without increasing the additional hardware costs such as acoustics or vision, it deeply analyzes the bidirectional physical mapping relationship between force and displacement through a deep learning network with a specific architecture.

[0023] The implementation process of the method described in this application will be described in detail below with reference to specific embodiments. It should be noted that this embodiment is only used to explain this application and is not intended to limit the scope of protection of this application. Conventional adjustments or substitutions of each step by those skilled in the art without departing from the concept of this application should be included in the scope of protection of this application.

[0024] Please see Figure 1 , Figure 1A flowchart illustrating a method for detecting press-fitting anomalies in a power transmission disc assembly according to some embodiments of the present disclosure is shown.

[0025] Step S1: Acquire the displacement and force sequences of the effective pressing segment and generate the corresponding force increment sequence. After the pressing process starts, the data acquisition card synchronously reads the analog signals from the piezoelectric sensor and the grating ruler at a set high-frequency sampling rate and converts them into digital signals. In one example, the system can be configured with a sampling rate of 5000 Hz to ensure that it can capture stick-slip friction abrupt changes lasting only a few milliseconds. The acquisition system monitors the spindle status in real time. When the read displacement value exceeds a certain proportion of the total stroke (e.g., 10%) and the synchronously read pressing force exceeds the set tooling contact preload, it is determined that the spindle has passed the no-load descent stage and entered the substantial interference fit range. At this time, the system starts to capture effective pressing segment data containing complete force characteristics from this node.

[0026] For example, the specific proportion of the total stroke is determined based on the proportion of the no-load stroke of different types of transmission disk assemblies. Specifically, the system monitors the first derivative of the displacement signal (i.e., instantaneous velocity) and the first derivative of the force signal in real time. When the first derivative of the displacement transitions from the negative peak region to the stable platform region and the first derivative of the force value continuously exceeds three times the standard deviation of the equipment noise floor threshold, it is determined that the spindle has entered the substantial interference fit range, and this moment is taken as the starting point of the effective pressing section.

[0027] Because the two types of sensors in industrial settings operate on different physical measurement principles, their electrical signal transmission and conversion circuits exhibit inherent delays at the microsecond level. Directly bundling data from the same timestamp would distort the mechanical curves at a microscopic scale. Therefore, before inputting the data into the model, the executing entity needs to extract the abrupt change points in the response slopes of the two types of sensors used for data acquisition during the initial macroscopic contact phase.

[0028] In practical applications, the executing entity extracts the array of original force values. and the original displacement array Take the first derivative with respect to time. and The time coordinates at which the derivatives of both signals first significantly exceed the equipment noise floor threshold are tracked. Corresponding to the instant the press-fit component first contacts the tooling, the timestamps of the abrupt changes in force signal response and displacement signal response are recorded. Subsequently, a time difference parameter is calculated based on these abrupt changes, reflecting the inherent phase shift between the two types of sensor channels. The equipment noise floor threshold is calculated by acquiring force signals from 1000 unloaded sampling points before press-fitting begins and calculating their standard deviation. Take later Sure.

[0029] After obtaining the above offset information, the initially acquired displacement data points are compensated for time axis using interpolation based on the time difference parameter. Taking linear interpolation as an example, the compensation calculation formula is set as follows:

[0030] In this formula, This represents the displacement at the current moment of the original data acquisition, with the dimension of millimeters. It represents the displacement at the previous sampling time, also in millimeters; and These represent the current and previous sampling times, respectively, with the unit being seconds; This refers to the time difference parameter calculated above, measured in seconds. The compensation operation is completed by traversing all acquisition nodes, and the system outputs the aligned displacement sequence.

[0031] In one embodiment, a one-dimensional array is constructed based on the aligned effective force value data to form a force value sequence. To highlight the high-frequency transient impact characteristics during the pressing process, the execution body performs a difference operation on the force value sequence along the time axis to calculate the force value difference between two adjacent sampling points, i.e.: This generates the corresponding force increment sequence.

[0032] For example, suppose that within a very short time window, the aligned displacement sequence contains data from five consecutive nodes: [12.010, 12.012, 12.014, 12.016, 12.018] mm; the corresponding force sequence contains five data points: [4000, 4100, 3200, 3800, 4050] N; the system performs backward differencing on the force sequence, generating the force increment sequence as follows: Newton. This operation transforms the hidden, dramatic fluctuations of the original absolute force value into a relative change with clearly defined positive and negative poles. Subsequently, the system concatenates the extracted displacement sequence, force value sequence, and force increment sequence along the feature channel dimension to construct a three-dimensional input tensor with a data format of (batch size, sequence length, number of feature channels), where the number of feature channels is fixed at 3.

[0033] Step S2: Input the displacement sequence, force value sequence, and force value increment sequence into the anomaly detection model for feature extraction to obtain the basic feature sequence.

[0034] In this embodiment, a trained anomaly detection model is pre-loaded into the memory of the edge computing node. See also... Figure 2 , Figure 2 A schematic diagram of an anomaly detection model structure according to some embodiments of the present disclosure is shown.

[0035] The overall architecture of the anomaly detection model includes a shallow feature extraction network, an attention extraction layer, a temporal aggregation network, and a fully connected classification network connected in sequence. The three-dimensional input tensor constructed in step S1 is directly fed into the shallow feature extraction network of the anomaly detection model.

[0036] The shallow feature extraction network consists of three stacked one-dimensional convolutional neural networks, specifically designed to capture the local waveform morphology and basic trend of the mechanical curve. Specifically, the first convolutional layer has 64 one-dimensional convolutional kernels of size 7 and a stride of 1. After extracting the input tensor, the network applies a modified linear unit activation function for nonlinear mapping. The second convolutional layer has 128 convolutional kernels of size 5 and a stride of 1, also connected to a modified linear unit. The third convolutional layer has 256 convolutional kernels of size 3 and a stride of 1, expanding the output channel dimension to 256. Through these three convolutional layers, the model initially deconstructs the basic physical information of the pressing process. Optionally, to reduce the inference latency of edge computing nodes and adapt to the long-range dependency characteristics of Bi-GRU, a max-pooling layer with a stride of 2 is connected after each one-dimensional convolutional layer in the shallow feature extraction network. Through three pooling operations, the original high-frequency sampled long sequence is downsampled and compressed by 8 times in the time dimension; the final output is data in tensor form (batch size, sequence length, 256), which is the basic feature sequence required for subsequent steps.

[0037] Step S3: Obtain the weighted feature sequence by performing operations through the attention extraction layer of the anomaly detection model. The attention extraction layer receives the basic feature sequence output by the previous module, and simultaneously receives the original displacement sequence, force value sequence, and force value increment sequence generated in step S1 as bypass inputs for computation, and executes the following internal sub-steps in sequence.

[0038] The attention extraction layer consists of a series of serial operator nodes, replacing the traditional neuron parameter layer. Specifically, it includes a trough localization node: an embedded local minimum search algorithm that scans the force value increment sequence and anchors the time point that produces a significant negative drop as the trough node.

[0039] Instantaneous stiffness inversion node: Performs difference quotient calculation on the input force and displacement sequences ( This generates a local stiffness variation curve parallel to the time axis.

[0040] Adaptive window boundary matcher: Based on the instantaneous stiffness curve and the set structural stiffness threshold, it backtracks to the left to locate the starting point and searches to the right to locate the ending point, thereby dynamically dividing the left and right time windows corresponding to the trough nodes.

[0041] The physical characteristic integral calculation graph further includes the following branches: a velocity compensation branch: calculating the quotient of instantaneous velocity and reference velocity to generate a dimensionless velocity compensation coefficient; a deformation index branch: performing the accumulation of the absolute values ​​of force / displacement change rates with velocity compensation in the left and right windows respectively, outputting the drop deformation energy index on the left and the recovery deformation energy index on the right; and a plastic work branch: performing the discrete integral accumulation of the product of average force and displacement increment in the entire calculation interval, outputting the irreversible plastic work characteristic representing energy dissipation.

[0042] Attention scalar generator: Input the plastic work feature into the exponential decay function to obtain the exponential function value, divide the right side and the left side deformation energy exponent to obtain the basic ratio, and finally fuse and output a dimensionless morphology comparison coefficient.

[0043] Specifically, Figure 3 A schematic diagram illustrating the execution flow of the attention extraction layer according to some embodiments of the present disclosure is shown. Figure 3 As shown, in step S301: locating the trough node in the force increment sequence where a sudden drop occurs. This step can be executed by the trough location node operator. The attention extraction layer first scans the force increment sequence in the input bypass. In actual press-fitting conditions, both normal stick-slip friction and abnormal microcrack propagation are accompanied by an instantaneous collapse of resistance. Therefore, the network traverses the force increment sequence through a local minimum search algorithm. When a negative force increment is detected at a sampling node, and its absolute value exceeds the preset background noise fluctuation range, the system determines that a force collapse has occurred at that location, thereby locating the trough node in the force increment sequence where a sudden drop occurs. Taking the numerical example given earlier, the value of the force increment sequence at the third node is -900 Newtons, which is anchored by the network as a key trough node.

[0044] Step S302: Determine the instantaneous stiffness sequence based on the force value sequence and the displacement sequence. This step can be executed by the instantaneous stiffness inversion node operator. To eliminate time-domain interference caused by press-fit speed fluctuations, the network abandons using a simple time span to measure abrupt changes and instead introduces material mechanical properties. The attention extraction layer retrieves the force value sequence and displacement sequence within the neighborhood of the trough node and performs difference quotient calculation point by point.

[0045] The specific calculation logic can be translated into the following formula:

[0046] In the formula, and These represent the force values ​​at the current and previous nodes, respectively. and These represent the displacements of the current and previous nodes, respectively, thereby determining the instantaneous stiffness sequence based on the force sequence and the displacement sequence. .

[0047] Assuming the force increment within the sampling interval is -900 Newtons and the corresponding displacement increment is 0.002 mm, the instantaneous stiffness calculation result for this interval is -45,000 Newtons per millimeter. By traversing all nodes in the neighborhood, the network constructs a local stiffness change curve parallel to the time axis in memory, with the dimension unified to Newtons per millimeter.

[0048] Step S303: Define the left and right time windows corresponding to the trough nodes based on the instantaneous stiffness sequence. This step can be performed by an adaptive window boundary matcher operator. After obtaining the instantaneous stiffness parameter, the network uses the physical boundary of this parameter to accurately divide the influence range of the abrupt event. The specific operation includes the following three judgment dimensions: First, the numerical states in the instantaneous stiffness sequence are traced back to the left along the time axis. Normal materials are in the elastic deformation stage during the initial pressing phase, and their stiffness is relatively stable; however, just before separation or fracture, the stiffness begins to deviate sharply. The network checks the stiffness of each node individually, defining the time node corresponding to the first stable value exceeding the set structural stiffness threshold as the starting point of the left-hand time window. The structural stiffness threshold refers to the theoretical normal rate of change of force for this type of transmission disk assembly at the corresponding pressing depth. Its value is usually pre-calibrated by process engineers using standard non-destructive testing samples, for example, calibrated to 800 Newtons per millimeter.

[0049] Secondly, the numerical state in the instantaneous stiffness sequence is searched to the right along the time axis. If it is stick-slip friction, the material will quickly re-engage and recover its stress after energy release; if it is a microcrack, structural damage will cause permanent stiffness loss. The network addresses to the right, defining the node whose value is restored to the structural stiffness threshold range and whose single-step displacement increment exceeds the minimum displacement setting value as the termination point of the right-hand time window. The minimum displacement setting value set here is used to filter the static level jitter of the sensor itself, ensuring that the pressing action is indeed continuing. It can be determined based on the static resolution and level noise amplitude of the grating ruler displacement sensor, for example, a value of 0.001 mm.

[0050] Furthermore, to prevent the model from getting stuck in an infinite rightward search loop when encountering a completely fractured workpiece, a fusing mechanism is introduced. This means that if the value fails to recover to the structural stiffness threshold within the set maximum search span, the cutoff point of that maximum search span is forcibly defined as the termination point. This maximum search span can be preset to 100 sampling periods in the parameter file. Through this dynamic search mechanism, the model defines the left and right time windows corresponding to the trough nodes based on the instantaneous stiffness sequence, achieving adaptive cutting of the boundary of abrupt events. The maximum search span... According to sampling frequency Determined by prior knowledge of the physical duration of the abnormal event. ,in The longest physical duration (in seconds) from occurrence to termination of a known stick-slip friction or microcrack fracture event is statistically derived from the destructive mechanics analysis of historical failed components; at a sampling rate of 5000 Hz, Taking 0.02s, therefore One sampling period.

[0051] In step S304: Based on the left and right time windows, extract the bidirectional deformation index and the irreversible plastic work characteristic reflecting the energy dissipation state. This step can be performed by the physical feature integral graph operator, aiming to quantify the degree of energy accumulation and release before and after the abrupt event. The bidirectional deformation index specifically includes the left drop deformation energy index and the right recovery deformation energy index. The specific operation for extracting the bidirectional deformation index includes performing a weighted integral of physical quantities on the left and right intervals.

[0052] When extracting features from the left half, the network obtains the absolute ratio between the force change and the corresponding displacement change of adjacent samples within the sampling interval of the left time window. Since the pressing speed of the press in industrial settings is not absolutely constant, speed fluctuations can cause strain rate effects, and directly accumulating the absolute ratios can lead to feature distortion. Therefore, speed compensation logic needs to be introduced.

[0053] The specific extraction logic for the speed compensation coefficient includes: the executing entity first calculates the instantaneous pressing displacement rate value of the corresponding stage by combining the displacement increment of the current sampling node with the sampling period. Next, select the baseline pressing speed value under the calibration conditions from the system configuration library. Here, represents the calibrated pressing speed of this type of transmission disc assembly as specified in the standard process card, with units of millimeters per second. The specific value is determined by process engineers based on the equipment's factory parameters and the average pressing speed of historical compliant batches. Finally, the speed compensation coefficient is calculated using a formula. :

[0054] Using the calculated The velocity compensation coefficient, which serves as the corresponding sampling interval, is used to calculate the left-side drop deformation energy index after obtaining this dimensionless coefficient. The specific formula is as follows:

[0055] Thus, the drop deformation energy index on the left side is obtained.

[0056] For example, assuming there are two valid sampling intervals within the left-hand time window, and the calculated absolute ratios are 8000 N / mm and 9000 N / mm respectively, and the current instantaneous press speed is relatively slow, the calculated speed compensation coefficient is 1.2. Then the drop deformation energy index on the left side... The calculation is as follows: That is, the cumulative total is Newtons per millimeter.

[0057] When extracting features from the right half, for the sampling interval within the right-side time window, since the oscillation after the abrupt change may contain minor secondary drops, the network first filters out the effective sampling intervals that produce positive restoring force values, i.e., it removes those noisy intervals where the force value continues to decrease. Subsequently, it calculates the basic ratio of the force value change to the corresponding displacement change in the effective sampling interval. Similarly, it calculates the right-side restoring deformation energy index. The specific formula is as follows:

[0058] Thus, the right-side recovery deformation energy index is obtained. Through this asymmetric conditional screening, the model can purely characterize the elastic recovery ability of the material.

[0059] To further distinguish between frictional dissipation and fracture dissipation from a thermodynamic perspective and extract irreversible plastic work characteristics that reflect the energy dissipation state, the specific operations include: extracting discrete sampling node pairs within the closed-loop calculation interval formed by the starting point and the ending point; establishing computational micro-elements through adjacent data points; and then obtaining the average force distribution between adjacent sampling node pairs, while simultaneously obtaining the corresponding displacement increment range.

[0060] The network uses the accumulated results of the obtained characteristics representing the dissipation of the hysteresis loop as the irreversible plastic work characteristic. The calculation formula is as follows:

[0061] In the formula, This represents the average force distribution (in Newtons). Given the displacement increment range (dimension in millimeters), the product of the two yields the approximate work area within this infinitesimal interval (dimension converted to millijoules), and then discrete integrals are performed on the combined area parameters.

[0062] Under normal stick-slip friction conditions, the calculation interval is extremely narrow and the force value rebounds rapidly, resulting in a very small cumulative value of irreversible plastic work. However, under microcrack initiation conditions, plastic yielding and fracture surface energy release occur inside the material, which not only lengthens the interval but also reconstructs the average force due to baseline collapse. The cumulative result will show an order-of-magnitude expansion.

[0063] In step S305: A morphology comparison coefficient is generated based on the deformation index and the irreversible plastic work feature. This coefficient is then used to weight the nodes of the basic feature sequence. This step can be executed by an attention scalar generator operator, which is responsible for fusing and mapping the previously acquired mechanical parameters into scalar weights recognizable by the neural network. Specifically, generating the morphology comparison coefficient based on the deformation index and the irreversible plastic work feature includes extracting the right-side recoverable deformation energy index. Relative to the drop deformation energy index on the left side The basic ratio result is calculated using the following formula: A relative proportion is established between the irreversible plastic work characteristic and a given reference plastic work constant. The reference plastic work constant is the maximum permissible frictional dissipation work experimentally calibrated for a specific transmission disc material; its dimension is millijoules, the same as the irreversible plastic work characteristic, thus making the relative proportion a dimensionless value. Using this relative proportion as the core independent variable, an exponential decay function transformation calculation incorporating a negative decay factor is performed to obtain the exponential function value.

[0064] In this process, the negative decay factor is defined as a dimensionless parameter that controls the steepness of the penalty curve. Its value is obtained through statistical calibration of a large number of failure samples. For example, a grid search is performed on the validation set with a step size of 0.1 in the range of 0.5 to 2.0, and the γ value that maximizes the classification accuracy of the validation set is selected as the final value. For example, it can be preset as an empirical parameter with a value range between 0.5 and 2.0. In this example, it is set to 1.0. The calculation method of the exponential decay function can be expressed as:

[0065] In the formula, Represents a negative decay factor. Represents the characteristics of irreversible plastic work. This represents the baseline plastic work constant. Using this exponential function, when the plastic work is exceptionally large, the obtained exponential function value will rapidly approach zero. Finally, the morphological comparison coefficients for generating dimensionless morphologies are obtained. .

[0066] After obtaining the morphology comparison coefficients, the network uses them as attention scalars and performs element-wise scalar multiplication with the basic feature sequence generated in step S2 within the local feature vector channel corresponding to the trough node. That is, the morphology comparison coefficients are used to weight the basic feature sequence at each node. When the physical phenomenon is stick-slip friction, the comparison coefficients are close to 1, and the local features are completely preserved and transmitted. When the physical phenomenon is microcracks, the comparison coefficients are severely attenuated, causing the network to forcibly suppress and filter out the original deceptive high-frequency features.

[0067] Step S4: Perform global temporal relationship aggregation on the weighted feature sequence of nodes using the anomaly detection model, and output the classification judgment result for controlling the pressing equipment. After the attention extraction layer completes its processing, it outputs a weighted and filtered 3D feature tensor, maintaining its shape as (batch size, sequence length, 256), which is then passed as input to the backend temporal aggregation network. This temporal aggregation network is constructed using two layers of bidirectional gated recurrent units. The network reads feature vectors along both the forward and reverse directions of the time series, and relies on internal update and reset gate mechanisms to contextually fuse the locally processed trough shape information with the global macroscopic push-in trend.

[0068] After bidirectional iterative traversal, the hidden state vector of the last time step is extracted and concatenated to form a dimensionality-reduced global representation vector. This global representation vector is fed into a fully connected classification network consisting of two linear mapping layers. The final linear layer is configured with three output neurons, corresponding to the three categories of labels: normal, microcrack anomaly, and dimensional anomaly. Finally, a normalization operation is performed using a flexible maximum function to generate a probability distribution vector of shape (batch size, 3). This probability distribution vector is the output of the model, i.e., the classification judgment result used to control the pressing equipment.

[0069] After receiving the classification result, when the probability of parsing the abnormal category exceeds the judgment threshold, the edge computing node immediately sends an interception command signal to the programmable logic controller to control the action of the robotic arm or waste sorting valve on the pressing line, directly removing the corresponding batch of unqualified power transmission disk assemblies from the main conveyor belt to prevent them from flowing into the next tooling testing stage.

[0070] In some embodiments, regarding model training, to ensure the generalization ability of the above-mentioned anomaly detection model under various complex batches, the model is trained with constrained parameters. The dataset required for model training comes from the historically accumulated sensor data of the entire pressing process. Its data format is consistent with the input of the previous steps and is accompanied by category labels marked by manual quality inspection.

[0071] Specifically, the dataset contains no fewer than 10,000 valid press-fit sample records, each record covering a complete displacement-force time sequence. Approximately 70% of the samples are normal, about 15% contain microcracks, and about 15% exhibit dimensional abnormalities. The samples originate from at least three different production batches and two press-fit speed settings (fast and slow). All sample category labels were manually verified and cross-validated through metallographic analysis and profilometry measurements.

[0072] After the third convolutional layer of the shallow feature extraction network, a batch normalization layer and a temporary fallback layer with a dropout rate of 0.25 are added; after the two bidirectional gated recurrent units of the temporal aggregation network, a temporary fallback layer with a dropout rate of 0.3 is added. During training, an early stopping mechanism is adopted: when the total error on the validation set does not show a decreasing trend after ten consecutive training epochs, training is terminated early and the network parameters are restored to the optimal epoch to prevent model overfitting.

[0073] The anomaly detection model performs parameter training and optimization based on an error function that incorporates physical consistency constraints. The corresponding error function construction logic includes the following two levels of calculation: In terms of classification, the network uses the probability distribution of input samples in the target classification state category, combined with cross-entropy, to calculate the baseline classification difference. This value measures the pure data-level deviation between the network's prediction and the actual human labels, and is a conventional dimensionless value.

[0074] In terms of physical boundaries, relying solely on data labels can easily lead to the model losing its physical common sense on edge test sets. Therefore, the executing entity extracts the final force parameter fed back at the end of the press-fit operation of the input sample. This parameter reflects the final clamping force level achieved by the interference fit, with the dimension of Newton. When the final force parameter is determined to be lower than the set minimum acceptable preload value, the formula for calculating the relative difference percentage is as follows:

[0075] In the formula, This is the lower limit of the minimum acceptable preload value. As the final force parameter, the model then generates a physical consistency deviation value. The calculation formula is as follows:

[0076] In the formula, The core mechanism of this design, which classifies the output as a "normal" state, is as follows: if the force value at the end of the pressing process is not up to standard, it indicates that the part is severely damaged. If the model happens to give a high "normal" prediction probability at this time, the product term will surge, severely penalizing the model. After completing the two parts of the calculation, the total error of updating the anomaly detection model network parameters is based on... The calculation formula is:

[0077] In the formula, This represents the numerical value of the categorical difference. To adjust the weights, a weighted summation is used as the basis for the total error. In one training parameter configuration example, a weight of 0.5 is added before the physical consistency deviation value during weighted summation. The optimizer uses the adaptive moment estimation optimization algorithm. The initial learning rate is configured to 0.001 and gradually decreases with each training round. The batch size is set to 64. The model undergoes two hundred rounds of forward propagation and backward propagation calculations until the total error converges to an equilibrium state, thus solidifying the internal convolutional kernels and connection weights.

[0078] In practical engineering deployments, the anomaly detection model can run on edge computing nodes. For example, this node uses an ARM Cortex-A53 quad-core processor with a clock speed of 1.5GHz, 2GB of DDR4 memory, and integrates an NPU acceleration unit (with a computing power of no less than 1 TOPS). Model inference employs an INT8 quantization deployment method, unifying the floating-point weight parameters of the shallow feature extraction network and the temporal aggregation network into 8-bit integers to reduce memory consumption and computational latency. Actual testing shows that at a sampling rate of 5000Hz, the average latency of a single inference is approximately 8.2ms, far less than the 200μs data accumulation time window corresponding to a single sampling period (i.e., inference is triggered after accumulating a batch of data, rather than triggering inference per sampling point), meeting the real-time pressing cycle requirements (typical pressing cycle time is 3 to 5 seconds per piece).

[0079] Regarding sensor installation, the piezoelectric force sensor is installed on the thrust transmission path of the pressing spindle, specifically between the end of the spindle and the flange connecting the pressure head; the grating ruler displacement sensor is installed parallel to the spindle axis on the side of the column guide rail, with its reading head rigidly connected to the moving parts of the spindle. After installation, a pneumatic operation is performed using a standard rigid block (without elastic deformation or interference fit). The inherent offset reference value is calibrated using the time difference between the force signal and the displacement signal response abrupt change point. In subsequent formal production, the inherent offset of the system is subtracted from this reference value to obtain the true relative delay caused by the physical characteristics of the workpiece.

[0080] Therefore, this embodiment seamlessly embeds instantaneous stiffness search, strain rate compensation deformation energy calculation, and irreversible plastic work integration mechanism into the attention mapping layer of a deep learning network, achieving decoupling of complex waveforms from the perspective of physical constitutive relations rather than simply data statistics. This effectively solves the problem of misjudging stick-slip friction fluctuations and microcrack fracture fluctuations in the automotive parts press-fitting industry. Without adding external hardware testing equipment, it eliminates time-domain and velocity-domain interference caused by flexible working conditions, significantly improving the accuracy of defective product interception and production first-pass yield.

[0081] Figure 4 A schematic diagram illustrating the structure of an electronic device according to some embodiments of the present disclosure is shown. For example... Figure 4As shown, the electronic device includes a memory 400 and a processor 402. The memory stores instructions that can implement the above-described combination of... Figure 1 The described method allows the electronic device to be an edge smart terminal deployed on the production line.

[0082] In the above embodiments, the descriptions of each embodiment have different focuses. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The above descriptions are merely preferred embodiments of this application and explanations of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by specific combinations of the above technical features, but should also cover other technical solutions formed by arbitrary combinations of the above technical features or their equivalent features without departing from the inventive concept.

Claims

1. A method for detecting abnormal pressing of a power transmission disc assembly, characterized in that, include: S1, obtain the displacement sequence and force value sequence of the effective section of the press fitting, and generate the corresponding force value increment sequence; S2, input the displacement sequence, force value sequence and force value increment sequence into the shallow feature extraction network of the anomaly detection model to extract features and obtain the basic feature sequence; S3, obtaining the weighted feature sequence by performing operations through the attention extraction layer of the anomaly detection model, specifically including: Locate the trough nodes in the force increment sequence that exhibit a sudden drop. The instantaneous stiffness sequence is determined based on the force value sequence and the displacement sequence, and the left and right time windows corresponding to the trough nodes are defined based on the instantaneous stiffness sequence. Based on the left and right time windows, a bidirectional deformation index is extracted, and an irreversible plastic work characteristic reflecting the energy dissipation state is extracted. Based on the deformation index and the irreversible plastic work feature, a morphology comparison coefficient is generated, and the morphology comparison coefficient is used to perform node weighting on the basic feature sequence; S4. The anomaly detection model is used to perform global temporal relationship aggregation on the weighted feature sequence of nodes, and output the classification judgment result for controlling the pressing equipment.

2. The method according to claim 1, characterized in that, Obtaining the displacement sequence includes: Extract the abrupt change points of the response slope of the two types of sensors used for data acquisition in the initial stage of macroscopic contact; calculate the time difference parameter based on the abrupt change points; perform time axis compensation operation on the initially acquired displacement data points using interpolation based on the time difference parameter, and output the aligned displacement sequence.

3. The method according to claim 1, characterized in that, The left and right time windows corresponding to the trough nodes are defined based on the instantaneous stiffness sequence, specifically including: Tracing back the numerical state of the instantaneous stiffness sequence to the left along the time axis, the time node corresponding to the first stable value exceeding the set structural stiffness threshold is defined as the starting point of the left time window; searching the numerical state of the instantaneous stiffness sequence to the right along the time axis, the node where the value recovers to the range of the structural stiffness threshold and the single-step displacement increment exceeds the minimum displacement setting value is defined as the ending point of the right time window; when the value fails to recover to the range of the structural stiffness threshold within the set maximum search span, the cutoff position of the maximum search span is forcibly defined as the ending point.

4. The method according to claim 3, characterized in that, The bidirectional deformation index includes the left drop deformation energy index and the right recovery deformation energy index; extracting the left drop deformation energy index specifically includes: For the sampling interval within the left time window, the absolute ratio between the force change and the corresponding displacement change of adjacent samples is obtained. The absolute ratio is multiplied by the velocity compensation coefficient and then accumulated through interval traversal to obtain the left drop deformation energy index. For the sampling interval within the right-hand time window, select the effective sampling interval that produces positive restoring force value, calculate the basic ratio of the force value change to the corresponding displacement change in the effective sampling interval, multiply the basic ratio by the velocity compensation coefficient, and then perform interval traversal and accumulation to obtain the right-hand restoring deformation energy index.

5. The method according to claim 4, characterized in that, Extracting the right-side recoverable deformation energy index specifically includes: For the sampling interval within the right-hand time window, select the effective sampling interval that produces positive restoring force value, calculate the basic ratio of the force value change to the corresponding displacement change in the effective sampling interval, multiply the basic ratio by the velocity compensation coefficient, and then perform interval traversal and accumulation to obtain the right-hand restoring deformation energy index.

6. The method according to claim 4, characterized in that, The extraction of the speed compensation coefficient specifically includes: Calculate the instantaneous pressing displacement rate value of the current corresponding stage; select the reference pressing speed value under the calibration condition; use the ratio of the reference pressing speed value to the instantaneous pressing displacement rate value as the speed compensation coefficient of the corresponding sampling interval.

7. The method according to claim 3, characterized in that, Extracting the irreversible plastic work characteristic reflecting the energy dissipation state specifically includes: extracting discrete sampling node pairs within a closed-loop calculation interval formed by the starting point and the ending point; obtaining the average force distribution between adjacent sampling node pairs and the corresponding displacement increment range; multiplying and combining the average force distribution with the corresponding displacement increment range; performing discrete integral accumulation on the combined area parameters; and using the accumulated result representing the hysteresis loop dissipation characteristic as the irreversible plastic work characteristic.

8. The method according to claim 4 or 7, characterized in that, Based on the deformation index and the irreversible plastic work characteristic, a morphology comparison coefficient is generated, specifically including: Extract the basic ratio of the right-side recovery deformation energy index to the left-side drop deformation energy index; establish the relative proportion between the irreversible plastic work characteristic and the given benchmark plastic work constant; use the relative proportion as the core independent variable to perform an exponential decay function transformation calculation combined with a negative decay factor to obtain the exponential function value; construct the dimensionless morphological comparison coefficient by combining the basic ratio and the exponential function value through a multiplicative combination relationship.

9. The method according to claim 1, characterized in that, The anomaly detection model is trained and optimized based on an error function that incorporates physical consistency constraints. The corresponding error function construction logic includes: using the probability distribution of the input sample in the target classification state category and combining it with cross-entropy to calculate the classification difference value of the benchmark; extracting the final force value parameter fed back by the input sample at the end of the pressing operation; when it is determined that the final force value parameter is lower than the set minimum qualified preload value lower limit, generating a physical penalty term based on the relative difference ratio of the final force value parameter deviating from the minimum qualified preload value lower limit; fusing the physical penalty term with the predicted probability value when the classification output is the normal state category to form a physical consistency deviation value; and weighting and summing the classification difference value and the physical consistency deviation value according to the weight allocation to serve as the total error basis for updating the network parameters of the anomaly detection model.

10. A device for detecting abnormal pressing of a power transmission disc assembly, characterized in that, include: It includes a memory and a processor, the memory being configured to store computer-executable instructions thereon, which, when executed on the processor, perform the method of any one of claims 1 to 9.

Images