Cloud-based Neural Network Intelligent Recognition System and Method for Sheet Metal Quality Diagnosis

Abstract

This invention relates to the field of computer system technology based on a specific computational model, and discloses a cloud-based neural network intelligent recognition system and method for sheet metal quality diagnosis. The system includes: a feature splitting module that splits sheet metal state data to a high-frequency transient and steady-state cumulative feature calculation module to produce feature matrices and tensors; a dynamic temporal hedging compensation module that schedules cascaded buffer operators to synchronously align features based on the asynchronous lag deviation to construct a fused feature space tensor; a state deviation sensing module that introduces an adaptive decay matrix to correct connection weights when the state singular value distribution entropy exceeds a preset threshold; and an adaptive gradient dynamic compensation module that updates the segmentation judgment parameters based on the classification probability tensor. This invention smooths out convergence damping oscillations of the parameter matrix, effectively reduces feature redundancy, shortens the number of network convergence rounds, and eliminates the defect diagnosis generalization dead zone under extreme working conditions.

Description

Technical Field

[0001] This invention relates to a cloud-based neural network intelligent recognition system and method for sheet metal quality diagnosis, belonging to the technical field of computer system technology with specific computational models. Background Technology

[0002] The current distributed computing array deployment computing model processes continuous state signals. It completes feature extraction by reconstructing the topology of the hidden layer computing graph, which constitutes the mainstream architecture of quality diagnostic technology. In this architecture, the model uses distributed computing power to complete multi-level feature mapping and transmits synaptic connection weight parameters in the propagation path according to the inter-layer state transition rules, providing signal classification output under stable operating conditions.

[0003] However, sheet metal materials accumulate nonlinear strain during stamping. When high-dimensional transient strain sequences flow into the feature topological space, the interlayer state transition matrix generates exponential multiplication operations over consecutive time steps, leading to cascading accumulation of temporal noise and gradient deadlock within the computational model. To maintain classification performance, traditional approaches involve horizontally expanding the neuron scale or utilizing idle computational cores within the system to construct post-processing verification units. These methods increase the parameter scale of the dense synapse matrix, resulting in memory preemption and cross-thread communication delays. When nonlinear deviations or other edge states occur, the asynchronous delay of the computational graph can easily cause the predicted classification probability to fall into the generalization dead zone. The aforementioned limitations in expanding hardware scale restrict the flexibility of the system architecture, and the supporting software control and parameter optimization methods also have shortcomings. For example, the Chinese invention with publication number CN117494569A... The patent application disclosed a method and device for optimizing the sheet metal thickness of electrical equipment skeletons. It uses the Latin hypercube method for experimental design and combines finite element simulation results to train a neural network to run a multi-objective optimization algorithm. This is an idealized design that relies on offline static simulation samples. This scheme implicitly depends on the underlying premise that the operating parameters are within a preset stable range. When faced with dynamic evolution scenarios such as nonlinear rheological changes in material stress and the influx of transient strain sequences during the on-site stamping process, the statically trained network model is fundamentally mismatched with the actual complex boundary conditions. It cannot suppress the cascading accumulation of temporal noise in continuous time steps. When encountering non-ideal operating conditions such as nonlinear deviations or cross-thread communication delays, it is easy to cause a dead zone in the generalization of defect diagnosis due to the divergence of network synaptic connection parameters. Achieving dynamic feedforward decoupling and temporal hedging of inter-layer state transitions in a distributed computing environment is a technical bottleneck that needs to be solved.

[0004] Therefore, how to construct interlayer synaptic connection control rules that combine feedforward decoupling and dynamic temporal hedging, and improve the convergence accuracy and adaptability of the model under non-stationary temporal characteristics, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A cloud-based neural network intelligent recognition system for sheet metal quality diagnosis, comprising:

[0006] The feature splitting module is used to acquire sheet metal state data and split the sheet metal state data to the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module, so as to run and process to produce transient feature mapping matrix and steady-state cumulative feature tensor respectively;

[0007] The dynamic timing offset compensation module, connected to the feature splitting module, is used to collect the timing asynchronous lag deviation between the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module during collaborative operation, and to schedule the dynamic cascaded buffer operator to offset the calculation time difference, so as to synchronously align the transient feature mapping matrix and the steady-state accumulated feature tensor and construct the fused feature space tensor.

[0008] The state deviation perception module, connected to the feature diversion module, is used to extract the inter-layer state transition matrix of the high-frequency transient feature calculation module and calculate the state singular value distribution entropy. When the state singular value distribution entropy is greater than the preset critical threshold, the normal forward state flow is blocked, and an adaptive decay matrix is ​​introduced to correct the inter-layer synaptic connection weight parameters.

[0009] The adaptive gradient dynamic compensation module is connected to the feature splitting module and the state deviation perception module respectively. It is used to calculate the evolution gradient of the prediction loss based on the generated classification probability tensor and update the splitting decision parameters of the feature splitting module accordingly.

[0010] Preferably, when calculating the singular value distribution entropy, the state deviation sensing module performs singular value spectral decomposition on the inter-layer state transition matrix to extract the discrete distribution characteristics of each singular value component, and calculates a scalar value characterizing the uniformity of the matrix spectrum as the singular value distribution entropy; when introducing the adaptive decay matrix, the state deviation sensing module indexes the corresponding discrete stepping coefficient in a preset memory address according to the magnitude of the singular value distribution entropy to construct a static adaptive decay matrix.

[0011] Preferably, when collecting the time-series asynchronous lag deviation, the dynamic time-series offset compensation module is used to quantitatively collect the time-series asynchronous lag deviation caused by the difference in topological complexity between the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module during parallel computation, and convert the time-series asynchronous lag deviation into a time-step delay deviation value; when scheduling the dynamic cascaded buffer operator, the dynamic time-series offset compensation module is used to schedule and call the dynamic cascaded buffer operator on the time axis to offset the computation time difference, so as to ensure that the time steps of the transient feature mapping matrix and the steady-state cumulative feature tensor are synchronously aligned.

[0012] Preferably, the adaptive gradient dynamic compensation module is used to collect the ratio of the classification probability tensor to the historical residual stored in the cloud and calculate the gradient of the predicted loss evolution. When the gradient of the predicted loss evolution monotonically increases within three consecutive stamping cycles, the adaptive gradient dynamic compensation module updates the segmentation judgment parameters of the feature splitting module to adjust the splitting ratio of input data to the high-frequency transient feature calculation module.

[0013] Preferably, the system also includes an adaptive degradation optimization module, which is connected to the state deviation perception module and is used to periodically monitor the communication delay parameters between the cloud and the terminal. When the communication delay parameters exceed the preset delay threshold of 12ms for three consecutive processing cycles, the dimension of the inter-layer state transition matrix is ​​reduced to reduce the feature extraction calculation frequency.

[0014] Preferably, the feature splitting module is also used to receive the original state sequence collected by the sheet metal stamping sensor and transmitted through the industrial network, and to perform normalization processing on the original state sequence to eliminate dimensional differences and generate sheet metal state data.

[0015] Preferably, the preset critical threshold set in the state deviation sensing module is 0.78. When the calculated state singular value distribution entropy is not greater than 0.78, the state deviation sensing module maintains the normal forward state flow and does not construct an adaptive decay matrix.

[0016] Preferably, the dynamic temporal hedging compensation module is also used to perform element-wise weighted fusion of the synchronously aligned transient feature mapping matrix and the steady-state cumulative feature tensor under its internal feature fusion configuration, so as to construct a fused feature space tensor to suppress the temporal misalignment of high-dimensional non-steady-state features during fusion.

[0017] Preferably, the system also includes a classification output module, which is connected to the dynamic temporal hedging compensation module. The classification output module is used to input the fused feature space tensor into the fully connected layer for probability calculation, produce a classification probability tensor that characterizes the risk of sheet metal stamping cracking, wrinkling and springback defects, and send the classification probability tensor to the cloud storage array.

[0018] A cloud-based neural network intelligent recognition method for sheet metal quality diagnosis, used to run a cloud-based neural network intelligent recognition system for sheet metal quality diagnosis, includes the following steps:

[0019] Step S1: Use the feature splitting module to obtain sheet metal state data, and split the sheet metal state data to the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module, so as to run the processing to produce transient feature mapping matrix and steady-state cumulative feature tensor respectively.

[0020] Step S2: The dynamic time-series offset compensation module is used to collect the time-series asynchronous lag deviation between the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module during collaborative operation, and the dynamic cascaded buffer operator is scheduled to offset the calculation time difference, so as to synchronously align the transient feature mapping matrix and the steady-state accumulated feature tensor and construct the fused feature space tensor.

[0021] Step S3: Use the state deviation perception module to extract the interlayer state transition matrix of the high-frequency transient feature calculation module and calculate the state singular value distribution entropy. When the state singular value distribution entropy is greater than the preset critical threshold, block the normal forward state flow and introduce an adaptive decay matrix to correct the interlayer synaptic connection weight parameters.

[0022] Step S4: The adaptive gradient dynamic compensation module calculates the predicted loss evolution gradient based on the generated classification probability tensor, and updates the segmentation decision parameters of the feature splitting module accordingly.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] 1. In cloud-based neural network intelligent recognition, the transient strain displacement time series data stream of the formed cross section is transformed and split through the feature splitting module. The first-order differential energy gradient splitting operator configured inside separates the high-dimensional state feature tensor into high-frequency variation components and low-frequency trend components, and transmits them hierarchically to the parallel-deployed high-frequency rheological synaptic response partition and steady-state feature evolution accumulation partition. This splitting and collaborative construction method blocks the serial transmission path of time series noise in a single graph topology, reduces the algebraic correlation of the hidden layer state matrix, prevents the parameter matrix from getting stuck in a local deadlock state during gradient optimization, smooths the dissipation resistance in the backpropagation process, and reduces the number of update rounds in the feature iteration convergence process.

[0025] 2. The dynamic temporal hedging compensation module quantitatively collects the temporal asynchronous lag deviation and time step offset caused by the difference in topological complexity between the high-frequency rheological synaptic response partition and the steady-state feature evolution accumulation partition during parallel computation. It schedules and calls the dynamic cascaded buffer operator on the time axis to offset the computational time difference, ensuring that the time steps of the transient feature mapping matrix and the steady-state cumulative feature tensor are synchronized and aligned. This solves the inherent time delay lag deviation of the model. In conjunction with the cross-regional cross-fusion flow architecture, it constructs a fused feature space tensor to avoid temporal misalignment and the risk of backward gradient explosion during the fusion of high-dimensional non-steady physical features.

[0026] 3. The state deviation sensing operator extracts the state transition matrix of the hidden layer of the high-frequency rheological synaptic response partition at the current time step. Singular value spectral decomposition is performed on this matrix to calculate the information entropy fluctuation rate, which characterizes the uniformity of the matrix spectrum. If this index is greater than the critical threshold of 0.78, the system truncates the linear state transfer of the forward propagation flow, activates the nonlinear attention enhancement matrix and injects it in situ as a multiplication weighting coefficient into the synaptic activation function of the corresponding partition, corrects the graph topology connection weight allocation parameters, reduces the sensitivity of the synapse to random industrial electromagnetic disturbances, and avoids the prediction probability generalization dead zone of the model parameters under the interference of highly coupled variation features. Attached Figure Description

[0027] Figure 1 This is a flowchart of the neural network processing method for sheet metal quality feature diversion and offset compensation according to the present invention;

[0028] Figure 2 This is a topology diagram of the intelligent identification system of the present invention, which includes production line nodes and cloud computing architecture.

[0029] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0030] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0031] A cloud-based neural network intelligent recognition system for sheet metal quality diagnosis includes:

[0032] The feature splitting module is used to acquire sheet metal state data and split the sheet metal state data to the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module, so as to run and process to produce transient feature mapping matrix and steady-state cumulative feature tensor respectively;

[0033] The dynamic timing offset compensation module, connected to the feature splitting module, is used to collect the timing asynchronous lag deviation between the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module during collaborative operation, and to schedule the dynamic cascaded buffer operator to offset the calculation time difference, so as to synchronously align the transient feature mapping matrix and the steady-state accumulated feature tensor and construct the fused feature space tensor.

[0034] The state deviation perception module, connected to the feature diversion module, is used to extract the inter-layer state transition matrix of the high-frequency transient feature calculation module and calculate the state singular value distribution entropy. When the state singular value distribution entropy is greater than the preset critical threshold, the normal forward state flow is blocked, and an adaptive decay matrix is ​​introduced to correct the inter-layer synaptic connection weight parameters.

[0035] The adaptive gradient dynamic compensation module is connected to the feature splitting module and the state deviation perception module respectively. It is used to calculate the evolution gradient of the prediction loss based on the generated classification probability tensor and update the splitting decision parameters of the feature splitting module accordingly.

[0036] Preferably, when calculating the singular value distribution entropy, the state deviation sensing module performs singular value spectral decomposition on the inter-layer state transition matrix to extract the discrete distribution characteristics of each singular value component, and calculates a scalar value characterizing the uniformity of the matrix spectrum as the singular value distribution entropy; when introducing the adaptive decay matrix, the state deviation sensing module indexes the corresponding discrete stepping coefficient in a preset memory address according to the magnitude of the singular value distribution entropy to construct a static adaptive decay matrix.

[0037] Preferably, when collecting the time-series asynchronous lag deviation, the dynamic time-series offset compensation module is used to quantitatively collect the time-series asynchronous lag deviation caused by the difference in topological complexity between the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module during parallel computation, and convert the time-series asynchronous lag deviation into a time-step delay deviation value; when scheduling the dynamic cascaded buffer operator, the dynamic time-series offset compensation module is used to schedule and call the dynamic cascaded buffer operator on the time axis to offset the computation time difference, so as to ensure that the time steps of the transient feature mapping matrix and the steady-state cumulative feature tensor are synchronously aligned.

[0038] Preferably, the adaptive gradient dynamic compensation module is used to collect the ratio of the classification probability tensor to the historical residual stored in the cloud and calculate the gradient of the predicted loss evolution. When the gradient of the predicted loss evolution monotonically increases within three consecutive stamping cycles, the adaptive gradient dynamic compensation module updates the segmentation judgment parameters of the feature splitting module to adjust the splitting ratio of input data to the high-frequency transient feature calculation module.

[0039] Preferably, the system also includes an adaptive degradation optimization module, which is connected to the state deviation perception module and is used to periodically monitor the communication delay parameters between the cloud and the terminal. When the communication delay parameters exceed the preset delay threshold of 12ms for three consecutive processing cycles, the dimension of the inter-layer state transition matrix is ​​reduced to reduce the feature extraction calculation frequency.

[0040] Preferably, the feature splitting module is also used to receive the original state sequence collected by the sheet metal stamping sensor and transmitted through the industrial network, and to perform normalization processing on the original state sequence to eliminate dimensional differences and generate sheet metal state data.

[0041] Preferably, the preset critical threshold set in the state deviation sensing module is 0.78. When the calculated state singular value distribution entropy is not greater than 0.78, the state deviation sensing module maintains the normal forward state flow and does not construct an adaptive decay matrix.

[0042] Preferably, the dynamic temporal hedging compensation module is also used to perform element-wise weighted fusion of the synchronously aligned transient feature mapping matrix and the steady-state cumulative feature tensor under its internal feature fusion configuration, so as to construct a fused feature space tensor to suppress the temporal misalignment of high-dimensional non-steady-state features during fusion.

[0043] Preferably, the system also includes a classification output module, which is connected to the dynamic temporal hedging compensation module. The classification output module is used to input the fused feature space tensor into the fully connected layer for probability calculation, produce a classification probability tensor that characterizes the risk of sheet metal stamping cracking, wrinkling and springback defects, and send the classification probability tensor to the cloud storage array.

[0044] A cloud-based neural network intelligent recognition method for sheet metal quality diagnosis, used to run a cloud-based neural network intelligent recognition system for sheet metal quality diagnosis, includes the following steps:

[0045] Step S1: Use the feature splitting module to obtain sheet metal state data, and split the sheet metal state data to the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module, so as to run the processing to produce transient feature mapping matrix and steady-state cumulative feature tensor respectively.

[0046] Step S2: The dynamic time-series offset compensation module is used to collect the time-series asynchronous lag deviation between the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module during collaborative operation, and the dynamic cascaded buffer operator is scheduled to offset the calculation time difference, so as to synchronously align the transient feature mapping matrix and the steady-state accumulated feature tensor and construct the fused feature space tensor.

[0047] Step S3: Use the state deviation perception module to extract the interlayer state transition matrix of the high-frequency transient feature calculation module and calculate the state singular value distribution entropy. When the state singular value distribution entropy is greater than the preset critical threshold, block the normal forward state flow and introduce an adaptive decay matrix to correct the interlayer synaptic connection weight parameters.

[0048] Step S4: The adaptive gradient dynamic compensation module calculates the predicted loss evolution gradient based on the generated classification probability tensor, and updates the segmentation decision parameters of the feature splitting module accordingly.

[0049] Example 1: In the monitoring scenario of an automotive body panel stamping production line, this system is used to diagnose the stress rebound deviation of the formed parts. The production line obtains the transient strain displacement time-series data stream of the formed section from the sensor array installed in the key pressing area of ​​the mold at a sampling frequency of 1kHz. This data stream contains the displacement change of continuous processing cycles. The feature splitting module receives the data stream, converts it into a high-dimensional state feature tensor, and stores it in the cache unit associated with the processor. The feature splitting module executes the first-order differential energy gradient splitting operator, that is, calculates the slope of the energy change of the tensor on the time axis. If the slope exceeds the preset energy gradient threshold, it is determined to be a high-frequency variation component; otherwise, it is classified as a low-frequency trend component. The dual-channel independent computation topology network receives the split data. The high-frequency rheological synaptic response partition consists of a set of neuron arrays with a weight update frequency of 200Hz, which is responsible for processing the high-frequency variation component; the steady-state feature evolution accumulation partition consists of a set of long short-term memory units, which process the low-frequency trend component.

[0050] The dynamic timing offset compensation module monitors the computation time of the two partitions and records the difference between them. If the high-frequency partition computation lags by 15ms, the module calculates the corresponding logical rack bounce and drives the cascaded buffer operator to perform a corresponding FIFO buffer shift on the time axis for the high-frequency partition output feature matrix. This ensures that the two feature streams are perfectly aligned in time at the convergence point of the cross-regional cross-fusion flow architecture. The state deviation sensing operator periodically extracts the hidden layer state transition matrix of the high-frequency partition. Then, perform singular value spectral decomposition on the matrix and calculate the singular value distribution entropy of the matrix. ,when When the critical threshold of 0.78 is exceeded, the operator activates the nonlinear attention enhancement matrix. By executing The Hadamard product operation corrects the topological connection weights, thereby suppressing surface noise interference. The adaptive gradient dynamic compensation loop continuously monitors the evolution of the prediction residuals of the classification probability tensor. If the residual ratio shows a monotonically increasing trend for three consecutive cycles, the discrimination boundary of the segmentation operator is adjusted according to the gradient to complete the closed-loop update of the parameters inside the system. This directly produces a classification probability tensor that characterizes the level of stress rebound geometric size deviation and feeds it back to the production line control terminal. When calculating the residual ratio, the loop extracts the multidimensional classification probability tensor output by the current processing cycle, performs a L2 norm modulus calculation on it to reduce its dimension to a one-dimensional scalar value, and uses this scalar value as the numerator. It then performs a homogeneous division operation with the scalar historical residual of the previous cycle stored in the cloud, thereby eliminating the mismatch between the dimensions and the unit of measurement and outputting an accurate scalar ratio that characterizes the evolution trend of the prediction loss.

[0051] Example 2: The current experiment was conducted on a monitoring platform based on an automotive body panel stamping production line. Data was collected using a vibration spectrum sensing module that is linked to the production line in real time. The module's frequency response range was set to 0.1kHz to 5kHz to ensure that it could cover the transient strain displacement signals generated by the forming section. To simulate real industrial electromagnetic background interference, Gaussian white noise with an amplitude of 0.5V and a signal-to-noise ratio of 20dB was superimposed at the input of the signal acquisition circuit to simulate the electromagnetic radiation interference generated by the motor and frequency converter on site. Simultaneously, 50Hz power frequency harmonic interference was superimposed. Cold-rolled low-carbon steel plate with a thickness of 1.2mm was selected as the test substrate. Continuous stamping operations were carried out under industrial conditions with a die pressing pressure set at 200 tons. To determine the optimal algorithm parameters, experimental group, control group A, and control group B were set up for comparative verification.

[0052] The experimental group ran the full-link intelligent recognition logic of this invention; control group A retained only the neural network inference function and removed the state deviation sensing module; control group B set the adjustment weight parameter in the adaptive gradient dynamic compensation loop to 0.12, which is outside the preferred value range of 0.15 to 0.35 determined by the invention. The test records showed that when the production line ran continuously for 1200 seconds, the temperature of the mold cavity rose to 85.5℃, causing the signal-to-noise ratio of the sensing signal to drop from 25.2dB to 17.8dB. Due to the lack of the state deviation sensing module, control group A failed to identify the change in input distribution, resulting in the uniformity of the singular value spectrum of the hidden layer state transition matrix. The output layer classification probability bias rate jumped sharply from 3.2% to 12.6% from 0.42 to 0.85, resulting in significant false alarms. Statistical analysis of experimental data from 5000 consecutive stamping cycles showed that when the singular value spectrum uniformity was below 0.78, the matrix spectrum distribution exhibited normal broadband characteristics. However, once it exceeded 0.78, it indicated an extreme clustering of singular values ​​in specific abnormal modes, suggesting that the system had entered a sudden environmental disturbance range. Therefore, 0.78 is the critical inflection point for determining whether the operating condition has deviated. Similarly, 1 The 2-millisecond communication latency threshold is set based on the maximum permissible transmission jitter limit of the high-frequency industrial network between the cloud and the terminal. If three consecutive cycles exceed 12 milliseconds, the transmission clock misalignment will cause the computational flow to diverge. Clearing synaptic weight coefficients with absolute values ​​below 0.05 to zero is to remove more than 45% of zero-approaching parameters from the dense matrix while maintaining the network classification accuracy from a precipitous drop. This allows the computational graph to be translated in situ into a discrete sparse matrix, thus achieving a sharp reduction in computational load. The experimental group system monitored... Once the value exceeds the trigger threshold of 0.78, the nonlinear attention enhancement matrix is ​​calculated. The system drives weight adjustment, stabilizing the deviation rate within 4.2%. Control group B, due to excessively low compensation parameters, failed to respond effectively to environmental disturbances, resulting in an output deviation rate as high as 15.8%, triggering the system's redundancy protection logic. Data analysis shows that the system logic response is best within the weight adjustment parameter range of 0.15 to 0.35. When the parameter is below 0.15, insufficient compensation weakens the system's suppression of power frequency interference; when the parameter is above 0.35, overfitting to high-frequency noise causes artificial oscillations in the feature space of the steady-state accumulated features, leading to the failure of the identification indicators. In this process, the aforementioned weight adjustment parameters serve as adaptive gradient dynamic compensation. The core scalar gain in the loop is pre-configured in the error feedback path of the adaptive gradient, specifically used to control the step size weight of the predicted loss evolution gradient on the modification of the segmentation decision parameter. This parameter is directly multiplied with the calculated loss slope in the algorithm loop to determine the fine adjustment amplitude of the segmentation decision boundary within a single processing cycle. When it is in the range of 0.15 to 0.35, the feedback convergence characteristic of the compensation loop is the most stable. Experimental results confirm that by using the state deviation sensing module to offset the feature deviation under non-stationary conditions in real time, the overall diagnostic accuracy of the system can be maintained above 95.5%, proving the synergistic effectiveness and feasibility of the combination of this technology in noisy environments.

[0053] Example 3: A quality diagnosis system for a current automotive body panel stamping production line includes a feature-based flow division module, a dynamic timing offset compensation module, a state deviation sensing module, and an adaptive gradient dynamic compensation module. In a specific stamping operation scenario, the contact pressure of the blank holder die and the sheet metal flow state are nonlinearly coupled. The system deploys a piezoelectric thin film sensor array on the blank holder surface of the die to acquire the strain displacement signal of the forming section in real time at a sampling frequency of 1kHz. This signal is converted into a digital feature matrix through a preprocessing circuit. The feature-based flow division module performs differential gradient calculation on the matrix to calculate the slope of energy change on the time axis. The part of the slope exceeding the preset gradient is classified as a high-frequency variation component, and the remaining part is regarded as a low-frequency trend component. This is to eliminate the mold... To mitigate aging drift caused by continuous operation, the system is set to trigger a background noise calibration every 500 stamped workpieces. During this calibration process, the system controls the upper die to perform an unloaded stroke, collects the background voltage value of the sensor array, and the processor calculates the statistical distribution of the voltage value, updates the mean and standard deviation, and writes them into the reference parameter register as the basis for zero-point drift correction in subsequent stamping cycles. When processing parallel partition features, the dynamic timing offset compensation module calculates the data processing time of each partition in real time. When the logical calculation lag of the high-frequency feature partition reaches 15ms, the system enables the FIFO cache shift operator to perform timing offset calibration on the high-frequency feature matrix on the time axis to ensure complete synchronization with the low-frequency steady-state tensor at the feature fusion point.

[0054] The state deviation perception module extracts the hidden layer state transition matrix of high-frequency feature partitions. Then, perform singular value spectral decomposition on the matrix and calculate the singular value distribution entropy of the matrix. The calculation formula is: ,in, The hidden layer state transition matrix The normalized singular values, when When the trigger threshold of 0.78 is exceeded, the system determines that the state evolution has entered the unsteady-state region. At this point, the state deviation perception module activates the nonlinear attention enhancement matrix. Drive the arithmetic unit to execute The Hadamard product operation reshapes the synaptic connection weights in the network topology in situ. When the prediction residual ratio of the classification probability tensor shows a monotonically increasing trend for three consecutive cycles, the adaptive gradient dynamic compensation module intervenes and adjusts the discrimination threshold of the splitting operator in reverse according to the residual gradient, thus completing the reset of the model convergence state. In the later stage of mold service, the system has a diagnostic accuracy of 94.2% for springback geometric dimension deviation, and the fluctuation range of classification probability residual is limited to within 0.05.

[0055] Example 4: In a manufacturing monitoring scenario for high-precision die stamping, the system is configured to diagnose the surface fatigue state of the die during the forming cycle using a neural network model. The production line continuously collects vibration and strain simulation signals installed on the base and support pillars of the stamping forming module at a sampling rate of 2kHz. The data transmission unit converts the signals into a time-based feature sequence. The feature splitting module receives and processes the feature sequence, using a first-order differential logic operator to calculate the feature intensity increment between adjacent time steps. ,in, This represents the feature intensity value at the current time step. The value represents the feature intensity from the previous time step. For the sampling period, when When the gradient is below the preset lower limit of 0.12, the splitting module classifies it into a steady-state feature accumulation calculation partition; otherwise, if... If the gradient is greater than the upper limit of 0.85, it is classified as a high-frequency transient feature mapping partition.

[0056] The state deviation sensing module calculates the tensor matrix in the partition of steady-state feature accumulation. Perform singular value spectral decomposition to obtain the eigenvalue sequence. Then calculate the state singular value distribution entropy. ,in, For normalized singular values, when When the value exceeds the set boundary of 0.65, the perception module determines that the calculation model has deviated from the initial healthy state of the mold and drives the weight correction logic. The logic module then calculates the feature deviation suppression operator. Perform Hadamard product operation to update the internal state transition matrix. To suppress the disordered expansion of feature distribution caused by surface fatigue, the dynamic timing offset compensation module monitors the logical timing stamps of each computing thread in the system in real time during feature fusion and calculates the inflow lag of features in each partition. If the time deviation of the feature flow at the merging point exceeds 10.2ms, the module calls the cascaded buffer operator to temporarily store the lagging feature matrix in the dynamic buffer pool and performs forced sequence shift according to the deviation. The processed feature tensor is assembled into a fused feature space matrix and sent to the subsequent recognition module to complete the mold fatigue state diagnosis. Experimental data shows that after introducing the state singular value distribution entropy control logic, the system's sensitivity to latent fatigue of the mold is improved by 18.5%, and the prediction deviation is stably maintained within 0.03 under long-term high load conditions.

[0057] Example 5: In a real-time diagnostic scenario of an automotive body panel stamping production line, the system is configured to diagnose the geometric deviation of stamping springback in the molded parts. The production line collects transient strain-displacement time-series data streams of the forming section from sensor arrays installed on the die blanking surface and key points of the forming cavity at a sampling frequency of 1kHz. The data receiving end acquires the time-series data stream and converts it into a floating-point time-series feature tensor containing 128 dimensions. The feature splitting module calls the first-order differential energy gradient splitting operator, using the formula... Calculate the current time step vector Compared with the vector of the previous time step The slope of the energy change, where, For the sampling period, when When the value exceeds 0.52, the shunt module marks the component as a high-frequency variation component; otherwise, it marks it as a low-frequency trend component and inputs it to the high-frequency rheological synaptic response partition and the steady-state feature evolution accumulation partition, respectively.

[0058] The high-frequency rheological synaptic response partitioning utilizes a computational model topology with specific synaptic connection weights to process high-frequency variational components. A state deviation sensing module extracts the partition's state at a specific time step in real time. State transition matrix and to Perform singular value spectral decomposition to obtain the sequence of singular value components. Through formula Calculate the state singular value distribution entropy ,in, For normalized singular values, when When the value exceeds 0.78, the discrete step coefficients corresponding to the state bias perception module index are used to construct a nonlinear attention enhancement matrix. and drive the arithmetic unit to execute With the next time step State transition matrixThe Hadamard product operation is used to update the synaptic connection weight parameters in the topology register of the hidden layer. Furthermore, the Hadamard product operation directly applies the coefficients of each point in the nonlinear attention enhancement matrix to the state transition vector at the current time step, thereby changing the overall flow attenuation rate of the state signal in the forward propagation graph topology. In the equivalent physical mapping of the neural network, this in-situ element-wise weighting of the inter-layer state transition matrix is ​​equivalent to applying a dynamic scaling regularization constraint to the inter-layer synaptic connection weight parameters without changing the physical values ​​of the hardware topology register. This mechanism blocks the abnormal forward flow caused by high-coupling variation features, achieving implicit adaptive dynamic correction of the synaptic connection weight allocation state. The state deviation sensing module retrieves the corresponding discrete stepping coefficient from the preset hexadecimal offset address in the static memory based on the magnitude of the distribution entropy of the state singular values ​​output by the high-frequency rheological synaptic response partition. The quantization optimization allocation path is as follows: when the distribution entropy is greater than 0.78 and not greater than 0.85, the cloud processor points the optimization pointer to the first register base address and reads the first discrete stepping coefficient 0.85. The first discrete stepping coefficient of 0.85 is multiplied at corresponding positions in the inter-layer state transition matrix. When the distribution entropy is greater than 0.85 but not greater than 0.92, the cloud processor points the optimization pointer to the base address of the second register and reads the second discrete stepping coefficient of 0.65. The second discrete stepping coefficient of 0.65 is then multiplied at corresponding positions in the inter-layer state transition matrix. When the distribution entropy is greater than 0.92, the cloud processor points the optimization pointer to the base address of the third register and reads the third discrete stepping coefficient of 0.40, reducing the forward signal propagation of the inter-layer state transition matrix. To suppress industrial electromagnetic radiation interference caused by the operation of frequency converter equipment at the stamping site, the dynamic timing offset compensation module collects the processing time difference between the high-frequency rheological synaptic response partition and the steady-state feature evolution accumulation partition in real time when processing partition features. When the endogenous time delay lag reaches 14.8ms, the module schedules the dynamic cascaded buffer operator to perform buffer translation on the high-frequency feature matrix on the time axis to ensure that the two feature flows are aligned in time steps at the convergence point of the cross-regional cross-fusion flow architecture. In actual operation, the module adjusts the sampling endogenous time delay by 14.8ms.Dividing 8 milliseconds by the sensor's signal sampling period of 1 millisecond yields a corresponding time-step delay deviation of 15 time steps, which is used as the translation step control command. Upon receiving this command, the dynamic cascaded buffer operator dynamically allocates a 32-block-deep, first-in-first-out dedicated queue buffer pool in memory, forcibly shifting the write pointer of the high-frequency feature matrix backward by 15 time addresses. This creates a physical delay equivalent to 15 milliseconds on the time axis for the high-frequency feature stream, achieving absolute time-step synchronization with the low-frequency steady-state tensor at the cross-regional hardware flow convergence point. The adaptive gradient dynamic compensation loop continuously monitors... The prediction residual evolution of the classification probability tensor is analyzed. If the residual ratio monotonically increases over three consecutive stamping cycles, the loop automatically updates the segmentation judgment parameters of the feature splitting module. Specifically, the loop calculates the mutual information divergence between the current cycle's classification probability tensor and the historical standard classification matrix in the cloud to obtain the prediction loss evolution gradient characterizing the classification dispersion. When this gradient is greater than 1.2 over three consecutive processing cycles, the judgment system experiences damped oscillation. At this point, the loop calls the gradient step size-based decreasing mapping logic to directly subtract the discrimination threshold of the currently stored first-order differential energy gradient segmentation operator from the original baseline of 0.52. Decrease by 0.05 until it reaches a hard lower limit of 0.35 to increase the proportion of input data diverted to the high-frequency transient feature calculation module; if the gradient continuously decreases, the discrimination threshold is increased in reverse with a fixed step size of 0.02, thereby achieving precise dynamic closed-loop adjustment of the segmentation judgment parameter. The adaptive gradient dynamic compensation module calculates the predicted loss evolution gradient based on the ratio of the collected classification probability tensor to the historical residual stored in the cloud. The closed-loop fine-tuning control path follows a preset decreasing mapping path, continuously comparing the L2 norm modulus scalar of the multidimensional classification probability tensor output in the current processing cycle with the historical residual scalar of the previous cycle, through... The homogeneous division operation outputs a scalar ratio that characterizes the trend of predicted loss evolution. When this ratio serves as the gradient for predicted loss evolution and increases monotonically in one direction over three consecutive stamping cycles, it indicates that the quality diagnostic model is experiencing damped oscillations. The preset gradient fine-tuning step size is then invoked, and the discrimination threshold of the currently stored first-order differential energy gradient splitting operator is reduced by 0.05 from the original baseline value of 0.52 until the discrimination threshold is reduced to the lower limit of 0.35. The damping of the parameter matrix is ​​corrected by increasing the proportion of input data diverted to the high-frequency transient feature calculation module. If the scalar ratio decreases continuously in subsequent cycles, it is set to 0.The fixed step size of 02 is increased in reverse to determine the threshold. The adaptive degradation optimization module starts the matrix pruning program when the communication delay parameter exceeds 12ms for three consecutive processing cycles. It maintains the physical topology connection of the current dynamic calculation graph and reduces the number of units actually participating in the general matrix multiplication calculation through the dynamic mask channel elimination method. The specific control procedure is as follows: decompose the singular values ​​of the current inter-layer state transition matrix of the high-frequency transient feature calculation module, extract and arrange all singular value components from largest to smallest, and calculate the proportion of the sum of the top k core singular value components to the total sum of all singular value components. When this proportion is greater than or equal to At 85%, the register values ​​corresponding to the remaining low-order singular components that are ranked lower are written to 0. Simultaneously, the synaptic weight coefficients in the inter-layer state transition matrix with absolute values ​​below 0.05 are forcibly cleared to zero, generating a discrete sparse matrix with the same physical dimension. This allows the cloud processor's graphics processing core to skip zero-value elements when reading the data bus, reducing the forward propagation hardware clock cycle and matching the current discrete network bandwidth. Diagnostic data records from the later stages of mold service show that the system's diagnostic accuracy for springback geometric dimension deviations is 94.8%, and the classification probability residual fluctuation range remains within 0.04.

[0059] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A cloud-based neural network intelligent recognition system for sheet metal quality diagnosis, characterized in that, include: The feature splitting module is used to acquire sheet metal state data and split the sheet metal state data to the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module, so as to run and process to produce transient feature mapping matrix and steady-state cumulative feature tensor respectively; The dynamic timing offset compensation module, connected to the feature splitting module, is used to collect the timing asynchronous lag deviation between the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module during collaborative operation, and to schedule the dynamic cascaded buffer operator to offset the calculation time difference, so as to synchronously align the transient feature mapping matrix and the steady-state accumulated feature tensor and construct the fused feature space tensor. The state deviation perception module, connected to the feature diversion module, is used to extract the inter-layer state transition matrix of the high-frequency transient feature calculation module and calculate the state singular value distribution entropy. When the state singular value distribution entropy is greater than the preset critical threshold, the normal forward state flow is blocked, and an adaptive decay matrix is ​​introduced to correct the inter-layer synaptic connection weight parameters. The adaptive gradient dynamic compensation module is connected to the feature splitting module and the state deviation perception module respectively. It is used to calculate the evolution gradient of the prediction loss based on the generated classification probability tensor and update the splitting decision parameters of the feature splitting module accordingly.

2. The cloud-based neural network intelligent recognition system for sheet metal quality diagnosis according to claim 1, characterized in that, When calculating the state singular value distribution entropy, the state deviation sensing module performs singular value spectral decomposition on the inter-layer state transition matrix to extract the discrete distribution characteristics of each singular value component, and calculates the scalar value characterizing the uniformity of the matrix spectrum as the state singular value distribution entropy. When introducing the adaptive decay matrix, the state deviation sensing module uses the indexing of the corresponding discrete stepping coefficients in a preset memory address based on the magnitude of the state singular value distribution entropy to construct a static adaptive decay matrix.

3. The cloud-based neural network intelligent recognition system for sheet metal quality diagnosis according to claim 1, characterized in that, When collecting the time-series asynchronous lag deviation, the dynamic timing hedging compensation module is used to quantitatively collect the time-series asynchronous lag deviation caused by the difference in topological complexity between the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module during parallel calculation, and convert the time-series asynchronous lag deviation into a time-step level delay deviation value. The dynamic timing offset compensation module is used to schedule and call the dynamic cascaded buffer operator on the time axis to offset the computation time difference, so as to ensure that the time steps of the transient feature mapping matrix and the steady-state cumulative feature tensor are synchronized and aligned.

4. The cloud-based neural network intelligent recognition system for sheet metal quality diagnosis according to claim 1, characterized in that, The adaptive gradient dynamic compensation module is used to collect the ratio of the classification probability tensor to the historical residual stored in the cloud and calculate the gradient of the predicted loss evolution. When the gradient of the predicted loss evolution monotonically increases within three consecutive stamping cycles, the adaptive gradient dynamic compensation module updates the segmentation judgment parameters of the feature splitting module to adjust the splitting ratio of input data to the high-frequency transient feature calculation module.

5. The cloud-based neural network intelligent recognition system for sheet metal quality diagnosis according to claim 1, characterized in that, The system also includes an adaptive degradation optimization module, which is connected to the state deviation perception module. It is used to periodically monitor the communication delay parameters between the cloud and the terminal. When the communication delay parameters exceed the preset delay threshold of 12ms for three consecutive processing cycles, the dimension of the inter-layer state transition matrix is ​​reduced to reduce the frequency of feature extraction calculation.

6. The cloud-based neural network intelligent recognition system for sheet metal quality diagnosis according to claim 1, characterized in that, The feature splitting module is also used to receive the raw state sequence collected by the sheet metal stamping sensor and transmit it through the industrial network, and to normalize the raw state sequence to eliminate dimensional differences and generate sheet metal state data.

7. A cloud-based neural network intelligent recognition system for sheet metal quality diagnosis according to claim 1, characterized in that, The preset critical threshold set in the state deviation sensing module is 0.

78. When the calculated state singular value distribution entropy is not greater than 0.78, the state deviation sensing module maintains the normal forward state flow and does not construct an adaptive decay matrix.

8. A cloud-based neural network intelligent recognition system for sheet metal quality diagnosis according to claim 1, characterized in that, The dynamic temporal hedging compensation module is also used to perform element-wise weighted fusion of the synchronously aligned transient feature mapping matrix and the steady-state cumulative feature tensor under its internal feature fusion configuration, so as to construct a fused feature space tensor to suppress the temporal misalignment of high-dimensional non-steady-state features during fusion.

9. A cloud-based neural network intelligent recognition system for sheet metal quality diagnosis according to claim 1, characterized in that, The system also includes a classification output module, which is connected to the dynamic temporal hedging compensation module. This module is used to input the fused feature space tensor into the fully connected layer for probability calculation, produce a classification probability tensor that characterizes the risk of sheet metal stamping cracking, wrinkling and springback defects, and send the classification probability tensor to the cloud storage array.

10. A cloud-based neural network intelligent recognition method for sheet metal quality diagnosis, used to run the cloud-based neural network intelligent recognition system for sheet metal quality diagnosis as described in claim 1, characterized in that, Includes the following steps: Step S1: Use the feature splitting module to obtain sheet metal state data, and split the sheet metal state data to the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module, so as to run the processing to produce transient feature mapping matrix and steady-state cumulative feature tensor respectively. Step S2: The dynamic time-series offset compensation module is used to collect the time-series asynchronous lag deviation between the high-frequency transient feature calculation module and the steady-state feature accumulation calculation module during collaborative operation, and the dynamic cascaded buffer operator is scheduled to offset the calculation time difference, so as to synchronously align the transient feature mapping matrix and the steady-state accumulated feature tensor and construct the fused feature space tensor. Step S3: Use the state deviation perception module to extract the interlayer state transition matrix of the high-frequency transient feature calculation module and calculate the state singular value distribution entropy. When the state singular value distribution entropy is greater than the preset critical threshold, block the normal forward state flow and introduce an adaptive decay matrix to correct the interlayer synaptic connection weight parameters. Step S4: The adaptive gradient dynamic compensation module calculates the predicted loss evolution gradient based on the generated classification probability tensor, and updates the segmentation decision parameters of the feature splitting module accordingly.

Images