Method and device for evaluating the quality of cold-formed steel roll forming

Abstract

The present invention provides a method and device for evaluating the forming quality of cold-bent steel rollers, which relates to the field of data processing technology. The typical frequency band characteristics of the monitoring data of the roller forming process are analyzed at multiple scales, and the corresponding joint characteristics are determined to perform quality evaluation. The transient impact characteristics of the roller vibration signal can be accurately captured in the time domain, and the main frequency component and its modulation sideband can be clearly separated in the frequency domain to adapt to the complex working conditions of the forming process. Under the premise of physical constraints on the nonlinear elastic characteristics of the roller, the non-steady-state characteristics corresponding to the roller quality degradation process are determined based on the mixed Gaussian distribution of the joint characteristics. It can combine the physical deformation of the roller to analyze and explain the internal mechanism of the complex system, capture the different distributions and corresponding nonlinear relationships of the multi-sensor time series data, and provide timely and accurate quality evaluation even if there are slight changes in the data or there are potential faults in the roller.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and device for evaluating the forming quality of cold-bent steel rolls. Background Art

[0002] Cold-formed steel rolls are crucial equipment in steel production and are widely used in the processing of cold-formed steel. Their forming quality directly impacts the precision and performance of the final product, making quality assessment of cold-formed steel rolls crucial.

[0003] However, traditional quality assessment methods mostly rely on manual experience or simple sensor monitoring. These methods are not only inefficient and inaccurate, but also difficult to adapt to the complex working conditions in the production process. Especially when the roller is in a non-steady-state working state, traditional methods cannot fully capture the subtle changes and potential failures in the complex quality degradation process, and thus cannot provide timely and accurate assessments. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a cold-bent steel roll forming quality assessment method and device, which can fully capture the subtle changes and potential failures in the complex quality degradation process and achieve timely and accurate quality assessment.

[0005] In a first aspect, an embodiment of the present invention provides a method for evaluating the forming quality of a cold-bent steel roll, wherein the method includes: monitoring the time series data of physical quantities of a preset roll during the forming process to obtain forming process monitoring data of the preset roll; the forming process monitoring data includes time series data corresponding to multiple sensors; using a preset wavelet packet basis function to determine the typical frequency band characteristics between the time series data of the forming process monitoring data, based on the typical frequency band characteristics, performing multi-scale convolution processing on the forming process monitoring data to determine the joint characteristics in the forming process monitoring data that characterize local mutations and global trends; according to a preset roll mechanics equation, performing physical constraint processing on the nonlinear elastic characteristics of the joint characteristics; and, based on the mixed Gaussian distribution of the joint characteristics, capturing the non-steady-state characteristics of the roll quality degradation process indicated by the joint characteristics under physical constraints; based on the non-steady-state characteristics, generating a quality category probability output result corresponding to the joint characteristics; and performing quality evaluation on the forming quality of the roll based on the quality category probability output result.

[0006] In combination with the first aspect, an embodiment of the present invention provides a first implementation method of the first aspect, wherein the step of using a preset wavelet packet basis function to determine the typical frequency band characteristics between the time series data of the molding process monitoring data includes: clustering the historical data spectrum of the molding process monitoring data, and taking the center of the cluster as the center frequency of the wavelet packet basis function; based on the center frequency, determining the typical frequency band characteristics corresponding to the molding process monitoring data.

[0007] In combination with the first aspect, an embodiment of the present invention provides a second implementation of the first aspect, wherein the step of monitoring the physical quantity time series data of the preset rolling mill during the forming process to obtain the forming process monitoring data of the preset rolling mill includes: using multiple sensors to collect data from the preset key positions of the preset rolling mill during the forming process, and obtaining the initial monitoring time series data of each sensor at the preset key position; maximizing the mutual information between the initial monitoring time series data of each sensor, and determining the lag time step corresponding to the initial monitoring time series data; determining the time series lag correlation corresponding to the initial monitoring time series data according to the lag time step; based on the time series lag correlation and the time window sliding variance of the initial monitoring time series data, the initial monitoring time series data is jointly normalized to determine the forming process monitoring data of the preset rolling mill.

[0008] In combination with the first aspect, an embodiment of the present invention provides a third implementation of the first aspect, wherein the step of physically constraining the nonlinear elastic characteristics of the joint feature according to the preset roller mechanics equation includes: performing singular value decomposition projection on the time window sliding average expectation of the joint feature to extract the main physical mode of the joint feature; determining the parameter constraint terms of the preset probabilistic neural network according to the preset roller mechanics equation; wherein the roller mechanics equation is calculated based on the material physical properties and material strain hardening characteristics of the roller; and determining the initialization weight parameters of the preset probabilistic neural network according to the parameter constraint terms and the main physical mode to physically constrain the nonlinear elastic characteristics of the joint feature based on the preset probabilistic neural network.

[0009] In combination with the first aspect, an embodiment of the present invention provides a fourth implementation method of the first aspect, wherein the step of capturing the non-steady-state characteristics of the roll quality degradation process indicated by the joint feature under physical constraints based on the mixed Gaussian distribution corresponding to the joint feature includes: determining the mixed Gaussian distribution components according to the typical degradation stage corresponding to the roll; determining the normal distribution corresponding to the joint feature based on the time step mean and time step standard deviation corresponding to the mixed Gaussian distribution components; based on the normal distribution, using a preset probabilistic neural network to capture the non-steady-state characteristics of the roll quality degradation process indicated by the joint feature; wherein the weight parameters of the probabilistic neural network are initialized based on the physical constraint terms corresponding to the roll mechanics equation.

[0010] In combination with the first aspect, an embodiment of the present invention provides a fifth implementation of the first aspect, wherein, based on the non-steady-state characteristics, the step of generating a quality category probability output result corresponding to the joint feature includes: obtaining the component contribution corresponding to the mixed Gaussian distribution component; mixing the normal distribution according to the component contribution and the last layer hidden state of the preset probabilistic neural network, and determining the quality category probability output result indicated by the joint feature.

[0011] In combination with the first aspect, an embodiment of the present invention provides a sixth implementation of the first aspect, wherein the weight parameters of the probabilistic neural network are determined based on preset adaptive momentum optimization parameters and adaptive learning rate; the adaptive momentum optimization parameters and / or adaptive learning rate are calculated based on the layer weight gradient of the probabilistic neural network; the method also includes: based on the preset roller mechanics equation to the layer weight sensitivity of the probabilistic neural network, physically perceived probabilistic discarding of the nodes of the probabilistic neural network; based on the current node of the probabilistic neural network and the loss function of the probabilistic neural network, determining the layer weight gradient of the probabilistic neural network.

[0012] In combination with the first aspect, an embodiment of the present invention provides a seventh implementation method of the first aspect, wherein the method for calculating the loss function includes: determining the distribution density penalty term corresponding to the joint feature based on the standard deviation of the mixed Gaussian distribution components; and determining the loss function of the probabilistic neural network based on the modified negative log-likelihood loss and the distribution density penalty term.

[0013] In combination with the first aspect, an embodiment of the present invention provides an eighth implementation of the first aspect, wherein the method for determining the adaptive learning rate includes: determining the feature distribution difference corresponding to the joint feature based on the kernel density estimation distribution of the joint feature corresponding to each iteration of the probabilistic neural network; and adaptively adjusting the learning rate of the probabilistic neural network based on the feature distribution difference.

[0014] In a second aspect, an embodiment of the present invention provides a cold-bent steel roller forming quality assessment device, wherein the device includes: a data monitoring module, which is used to monitor the physical quantity time series data of a preset roller during the forming process to obtain the forming process monitoring data of the preset roller; the forming process monitoring data includes time series data corresponding to multiple sensors; a feature extraction module, which is used to use a preset wavelet packet basis function to determine the typical frequency band characteristics between the time series data of the forming process monitoring data, and based on the typical frequency band characteristics, perform multi-scale convolution processing on the forming process monitoring data to determine the joint characteristics representing local mutations and global trends in the forming process monitoring data; a data processing module, which is used to perform physical constraint processing on the nonlinear elastic characteristics of the joint characteristics according to the preset roller mechanics equation; and, based on the mixed Gaussian distribution of the joint characteristics, capture the non-steady-state characteristics of the roller quality degradation process indicated by the joint characteristics under physical constraints; an execution module, which is used to generate a quality category probability output result corresponding to the joint characteristics based on the non-steady-state characteristics; and an output module, which is used to perform quality assessment on the forming quality of the roller based on the quality category probability output result.

[0015] The embodiments of the present invention bring the following beneficial effects: The embodiments of the present invention provide a method and device for evaluating the forming quality of cold-bent steel rolls. After analyzing the typical frequency band characteristics of the roll forming process monitoring data at multiple scales using wavelet packet basis functions, the corresponding joint characteristics are determined. The dynamic response behavior of different frequency components during the rolling process can be reflected based on the typical frequency band characteristics, thereby not only being able to characterize global trends (such as the degree of wear), but also to capture local mutations (such as faults or damage) represented by the data. Furthermore, combined with physical constraints on the nonlinear elastic characteristics of the rolls, the non-steady-state characteristics corresponding to the roll quality degradation process are determined based on the mixed Gaussian distribution of the joint characteristics. On the basis of ensuring that the characteristic changes conform to the actual force and deformation mechanism of the rolls, the different distributions and corresponding nonlinear relationships of the multi-sensor time series data can be captured by analyzing and interpreting the internal mechanisms of the complex system. Even if there are slight changes in the data or there are potential faults in the rolls, timely and accurate quality assessments can be provided.

[0016] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or understood by practicing the present invention. The purposes and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0017] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 A flow chart of a method for evaluating the quality of cold-bent steel roll forming provided by an embodiment of the present invention;

[0020] Figure 2 A flow chart of another method for evaluating the quality of cold-bent steel roll forming provided by an embodiment of the present invention;

[0021] Figure 3 A schematic diagram of a performance difference comparison result provided by an embodiment of the present invention;

[0022] Figure 4 A schematic diagram showing comparison results of feature fusion effectiveness corresponding to a normalization method provided in an embodiment of the present invention;

[0023] Figure 5 A schematic diagram of a parameter space distribution aggregation form provided by an embodiment of the present invention;

[0024] Figure 6 A schematic structural diagram of a cold-bent steel roll forming quality assessment device provided by an embodiment of the present invention;

[0025] Figure 7 A schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work shall fall within the scope of protection of the present invention.

[0027] In order to solve the above technical problems, the embodiments of the present invention provide a cold-bent steel roll forming quality assessment method and device, which can fully capture subtle changes and potential failures in the complex quality degradation process, and achieve timely and accurate quality assessment.

[0028] To facilitate understanding of this embodiment, a cold-bent steel roll forming quality assessment method disclosed in an embodiment of the present invention is first introduced in detail. Figure 1 A flow chart of a method for evaluating the quality of cold-bent steel roll forming provided by an embodiment of the present invention is shown, the method comprising:

[0029] Step S102 , monitoring the physical quantity time series data of the preset roller during the forming process to obtain the forming process monitoring data of the preset roller.

[0030] To promptly and accurately assess the quality of cold-formed steel roll forming, embodiments of the present invention collect monitoring data from various sensors during the roll forming process. These sensors include: a temperature sensor for monitoring the temperature of the rolls and the processing area; a pressure sensor for monitoring pressure changes during the roll forming process; a vibration sensor for detecting roll vibration and capturing roll deformation information; an acceleration sensor for monitoring roll acceleration and reflecting roll motion characteristics; and a strain sensor for monitoring the strain of the roll material to infer material deformation and performance changes. These sensor data are all time-series data. Multiple sensors installed at key locations throughout the roll forming process can collect relevant physical quantities (such as temperature, pressure, vibration, acceleration, and strain) in real time to generate corresponding forming process monitoring data. Furthermore, the collected data can be transmitted to a data processing system via Industrial Internet of Things technology for storage and analysis.

[0031] In step S104, a preset wavelet packet basis function is used to determine the typical frequency band characteristics between the time series data of the molding process monitoring data. Based on the typical frequency band characteristics, multi-scale convolution processing is performed on the molding process monitoring data to determine the joint characteristics of the molding process monitoring data that characterize local mutations and global trends.

[0032] Step S106, performing physical constraint processing on the nonlinear elastic characteristics of the joint feature according to the preset roll mechanics equation; and capturing the non-steady-state characteristics of the roll quality degradation process indicated by the joint feature under physical constraints based on the mixed Gaussian distribution of the joint feature.

[0033] Step S108 : generating a quality category probability output result corresponding to the joint feature based on the non-steady-state characteristics.

[0034] Step S110 : performing quality assessment on the forming quality of the roller based on the quality category probability output result.

[0035] Sensor monitoring data can be time-series data of physical quantities such as temperature, pressure, vibration, acceleration, and strain. It originates from sensors installed at key locations during the cold-bent steel roll forming process and is used to monitor the roll's operating status and forming quality in real time. This sensor monitoring data is multidimensional, time-series, subject to noise interference, and exhibits dynamic correlation differences between sensors. To eliminate dimensional differences and preserve multi-sensor synergy, embodiments of the present invention extract typical frequency band features from the data to determine joint features corresponding to multiple sensors. Furthermore, the non-steady-state characteristics within the joint features are captured to assess the forming quality of the target roll during the forming process based on these non-steady-state characteristics. In one embodiment, a threshold judgment method can be used to determine the classification results indicated by these non-steady-state characteristics, and a fuzzy logic strategy can also be used to determine the corresponding quality category probabilistic output results. Classification categories can include: normal state, slight deformation, moderate wear, severe fault, and critical failure.

[0036] In specific implementation, the embodiment of the present invention uses wavelet packet basis functions to analyze the typical frequency band characteristics of the roll forming process monitoring data at multiple scales, and then determines the corresponding joint characteristics. Based on the typical frequency band characteristics, the dynamic response behavior of different frequency components in the rolling process can be reflected, thereby not only being able to characterize the global trend (such as the degree of wear), but also to capture local mutations (such as faults or damage) represented by the data. Furthermore, combined with the physical constraints on the nonlinear elastic characteristics of the roll, the non-steady-state characteristics corresponding to the roll quality degradation process are determined based on the mixed Gaussian distribution of the joint characteristics. On the basis of ensuring that the characteristic changes are consistent with the actual force and deformation mechanism of the roll, the different distributions and corresponding nonlinear relationships of the multi-sensor time series data can be captured by analyzing and interpreting the internal mechanisms of the complex system. Even if there are slight changes in the data or there is a potential fault in the roll, the embodiment of the present invention can provide timely and accurate quality assessment.

[0037] Furthermore, based on the above embodiment, the embodiment of the present invention also provides another cold-bent steel roll forming quality assessment method. The embodiment of the present invention mainly describes the above steps S104-S106. Figure 2 A flow chart of another cold-bent steel roll forming quality assessment method provided by an embodiment of the present invention is shown. Figure 2 , the method comprises the following steps:

[0038] In step S202, a plurality of sensors are used to collect data from preset key positions of the preset rollers during the forming process, and initial monitoring time series data of each sensor at the preset key positions are obtained.

[0039] Based on the above-mentioned embodiments, the present invention utilizes multiple sensors to monitor the rolls and obtain forming process monitoring data. In a cold-formed steel roll forming quality assessment system, multiple sensors (such as force, displacement, temperature, and vibration) synchronously collect their respective operating status data. This data has inconsistent dimensions, and physical propagation delays or differences in response characteristics exist between the multiple sensors, resulting in temporal asynchrony in the collected data. Furthermore, the forming process of the rolls exhibits localized mutations and non-stationary fluctuations. Directly using raw data for modeling or evaluation can easily lead to model bias or misjudgment.

[0040] In order to solve the above problems, the embodiment of the present invention normalizes the data by taking into account the signal fluctuation characteristics and the coupling relationship between sensors. While eliminating the dimensional differences of sensors, it explicitly models the physical correlation across sensors in the cold-bend forming process, avoids the feature fragmentation of multi-sensor data, and enhances the consistency of cross-sensor features. In specific implementation, the embodiment of the present invention monitors the preset key positions based on each sensor to obtain the corresponding initial monitoring time series data. The initial monitoring time series data is processed through the following steps S204-S208, and the joint normalization strategy guided by the spatiotemporal covariance matrix is ​​used to avoid the independence of the dimensions corresponding to each sensor. The data is normalized based on the spatiotemporal coupling relationship between sensors to obtain the molding process monitoring data with effective feature fusion. Specifically, refer to the following steps S204-S208.

[0041] Step S204 , maximizing the mutual information between the initial monitoring time series data of each sensor, and determining the lag time step corresponding to the initial monitoring time series data.

[0042] Step S206: Determine the time series lag correlation corresponding to the initial monitoring time series data according to the lag time step.

[0043] The embodiment of the present invention optimizes the correlation of data by combining the mutual information between the collected data, adaptively determines the lag time step by maximizing the mutual information between the sensor data, and selects the time lag value that maximizes the mutual information between the sensor data. , capturing the mechanical response delay during roll deformation, such as the roll pressure change lags behind the vibration signal.

[0044] The embodiment of the present invention determines the dynamic correlation coefficient corresponding to each sensor based on the above-mentioned lag time step to reflect the cross-sensor timing lag correlation, such as the physical coupling between vibration and acceleration sensors. Specifically, the calculation method is as follows:

[0045]

[0046] is the data of the sth sensor at the t-Δ time step; is the dynamic correlation coefficient between the sth sensor and the s'th sensor. The dynamic correlation coefficient is used to quantify the degree of linear correlation between the time series data of two sensors at a specific time lag to capture the physical coupling relationship between sensors during the roll forming process, such as the delayed correlation between the vibration signal and the acceleration response.

[0047] In this formula, the numerator is the calculated covariance, which represents the difference between the current data change of the sth sensor and the data of the s'th sensor. The denominator is the product of the calculated standard deviations, where Part is the standard deviation of the sth sensor in the time window T, which measures its own fluctuation amplitude). Part is The sensor is lagging The standard deviation after the measurement is used to measure the fluctuation amplitude of the lagged data. The product of the two is used as the normalization factor to eliminate the dimension difference of the sensor.

[0048] In the calculation formula of the dynamic correlation coefficient, the calculation result is a correlation coefficient between [-1,1], that is: When , it indicates a strong positive correlation, such as when the vibration increases, the acceleration increases synchronously; When , it represents a strong negative correlation, such as the material strain decreases when the temperature increases; The dynamic correlation coefficient explicitly models the mechanical response delay in the roll deformation and is adaptively selected by maximizing the mutual information. , which ensures that the real physical coupling mechanism is captured.

[0049] Step S208: Based on the time series lag correlation and the time window sliding variance of the initial monitoring time series data, the initial monitoring time series data is jointly normalized to determine the forming process monitoring data of the preset rolling mill.

[0050] The sliding window method is used for each sensor sequence to calculate its local fluctuation intensity, thereby capturing the non-steady-state characteristics and abnormal fluctuations in the signal and enhancing the sensitivity to state changes. The calculation method is expressed as:

[0051]

[0052] Where, is the sliding variance, and T is the time window length, such as 48 sample values.

[0053] Furthermore, by combining the above sliding variance and time series lag correlation, the data collected by multiple sensors are jointly normalized to improve the physical consistency and dynamic adaptability of data fusion. Specifically, the normalization steps are as follows:

[0054]

[0055] Where, is the normalized output of the sth sensor at the tth time step; is the data of the sth sensor at the tth time step; is the sliding mean of the sth sensor data; For the The sliding mean of sensor data.

[0056] Referring to the above formula, the embodiments of the present invention use sliding variance to reflect the local volatility of the signal. Greater fluctuations result in smaller normalized values, thus suppressing anomalies. Furthermore, time-lag correlation is combined to reflect the degree of deviation from correlation with other sensors. Lower correlations result in greater penalties, enhancing global coordination. The normalization method of the embodiments of the present invention not only balances local stability and global consistency but also enhances the ability to identify non-stationary signals based on sliding variance, resulting in more accurate evaluation results.

[0057] Step S210 , clustering the historical data frequency spectrum of the molding process monitoring data, and taking the center of the cluster as the center frequency of the wavelet packet basis function.

[0058] Step S212: determining typical frequency band characteristics corresponding to the molding process monitoring data based on the center frequency.

[0059] Sensor data often contains a mixture of local mutations and global trends. To improve sensitivity to roll quality degradation patterns, an embodiment of the present invention adaptively distinguishes features from different frequency bands and, based on typical frequency band features, dynamically extracts features from the data (e.g., convolution processing) to effectively distinguish local mutations from global trends.

[0060] In a specific implementation, an embodiment of the present invention combines the signal decomposition of the wavelet packet transform to determine the typical frequency band characteristics. The wavelet packet transform can decompose the low-frequency part of the signal and can further subdivide the high-frequency part. In addition, its basic function (i.e., wavelet) can adjust the scale and position and has good time-frequency localization capabilities. In a specific implementation, the center frequency of the wavelet packet basis function is adaptively determined through spectral clustering, which can reflect the main frequency components under actual working conditions, that is, the typical frequency band characteristics (such as high-frequency noise and low-frequency deformation trends of roller vibration), making feature extraction more targeted. Among them, k-means clustering can be performed on the historical data spectrum, and the cluster center can be taken as the center frequency of the basis function.

[0061] In step S214 , based on the typical frequency band features, multi-scale convolution processing is performed on the molding process monitoring data to determine the joint features representing the local mutation and the global trend in the molding process monitoring data.

[0062] The embodiment of the present invention combines typical frequency band features for feature extraction, effectively distinguishes local mutations from global trends, and improves sensitivity to roll quality degradation patterns. Among them, the target feature in the forming process monitoring data is determined by multi-scale convolution processing. In the embodiment of the present invention, the target feature is a joint feature that characterizes local mutations and global trends. The embodiment of the present invention adopts a multi-scale adaptive convolution kernel module, combined with the above-mentioned wavelet packet decomposition and dynamic convolution weights, to achieve feature fusion while separating local mutations (such as roll jamming) and global trends (such as deformation accumulation caused by material fatigue) in the roll signal, thereby improving sensitivity to quality degradation patterns.

[0063] In the specific implementation, refer to the following formula to extract the corresponding features to capture the target features:

[0064]

[0065] Where, is the kth feature channel at time Output; is the total number of sensors; The convolution kernel time span is adaptively adjusted according to the timing characteristics and frequency domain characteristics of the data to ensure that the complete roll deformation cycle is covered, such as the vibration signal window of one roll rotation. Setting it to 0.5 times the sampling frequency (rounded up) makes it easier to capture the complete signal mutation window. For the Wavelet packet basis functions of the channel; is the convolution kernel weight between the kth feature channel and the sth sensor, and the calculation method is expressed as ; is a convolution kernel function, such as a Gaussian kernel function. It is a dynamic attention weight to achieve adaptive focusing across sensors and time series, such as the sudden change signal of the pressure sensor when the roller is overloaded. It can be seen from the above formula that the embodiment of the present invention determines the wavelet packet basis function in The typical frequency band features corresponding to the data at each moment are combined with the convolution kernel weights and dynamic attention weights to determine the corresponding target features.

[0066] In one embodiment, the calculation method of dynamic attention weight is expressed as:

[0067]

[0068] Referring to the above calculation formula, the embodiment of the present invention uses the hyperbolic tangent function to perform nonlinear compression and directional activation processing on the splicing vector of the sensor time series features and the hidden state, and calculates the dynamic attention weight of the probabilistic neural network based on the Softmax function.

[0069] in, is the Softmax function; It is a temperature parameter used to adjust the discreteness of attention weights. When its value is less than 1, the Softmax output is smoother, making the model pay equal attention to multi-sensor features. When its value is greater than 1, it strengthens significant features, which is suitable for capturing sudden failures of rollers. Preferably, Set to 0.5. It is a hyperbolic tangent function, which is used to perform nonlinear transformation on the concatenation vector of the sensor time series features and the hidden state. The nonlinear transformation can prevent numerical explosion while retaining the negative correlation. is the weight matrix of the probabilistic neural network; Indicates the The concatenation of the data of the kth sensor at time t-τ and the hidden state of the kth feature channel at time t-1; For the The raw data of each sensor at time t-τ; For the The hidden state of the feature channel at time t-1.

[0070] Further, Figure 3 A schematic diagram of performance difference comparison results corresponding to the embodiments of the present invention is provided. Figure 3 The performance comparison results of the traditional sliding average feature extraction implemented by the time-frequency joint analysis method and the multi-scale adaptive convolution kernel of the embodiment of the present invention are shown. Figure 3 It can be seen that the traditional method has obvious frequency aliasing phenomenon in the time-frequency spectrum, and cannot effectively distinguish between high-frequency impact components and low-frequency modulation signals, resulting in the loss of key process features; while the embodiment of the present invention, through the synergistic effect of wavelet packet decomposition and dynamic convolution weights, can accurately capture the transient impact characteristics of the roller vibration signal in the time domain, and clearly separate the main frequency component and its modulation sideband in the frequency domain, proving that the multi-scale convolution kernel has the ability to adaptively distinguish between local mutations and global trends of the signal, significantly enhancing the adaptability of the feature extraction process to complex working conditions.

[0071] Step S216: Physically constrain the nonlinear elastic characteristics of the joint feature according to the preset roller mechanics equation.

[0072] The above steps can effectively distinguish between local mutations and global trends in the collected data while eliminating the dimensional differences between multiple sensors and explicitly modeling the spatiotemporal coupling relationship between sensor data, so as to effectively capture the quality degradation pattern of the roller. As the equipment runs longer, the roller will degrade due to wear, fatigue, microcrack expansion, and other reasons. Correspondingly, the embodiment of the present invention also constrains the model based on the nonlinear elastic properties of the material to ensure that the processing of the model can comply with the physical laws of the roller forming process, thereby not only improving the generalization, but also capturing the non-steady-state characteristics in the "joint features" under the corresponding constraints. The non-steady-state characteristics are then classified and predicted, which can effectively monitor and judge the operating performance of the production process in a timely and accurate manner.

[0073] In this embodiment of the present invention, by determining the primary physical pattern between sensor features represented by the joint feature and combining it with the material mechanical parameters of the roller, parameter constraints are imposed on the model used in this embodiment of the present invention. This allows the model to identify target features (such as non-steady-state characteristics) based on the nonlinear elastic properties of the joint feature. In specific implementation, physical constraints are imposed through the following steps:

[0074] 1) Perform singular value decomposition projection on the time window sliding average expectation of the joint features to extract the main physical mode of the joint features.

[0075] First, the window is sliding based on the feature sequence of the joint feature to determine the sliding average expectation of the joint feature to smooth the noise, extract the local trend, and enhance the feature stability. The moving average expectation Perform singular value decomposition, take the first r principal component projections, and determine the singular value decomposition projection of the corresponding feature distribution . Among them, the principal component (Principal Component, PC) of the principal component projection can be used to characterize the main physical mode of the sensor feature, thereby extracting the main physical mode that determines the sensor feature. The main physical mode can represent the main structural direction in the feature space. For example, the principal component direction of pressure and vibration can avoid initialization from deviating from the actual working condition of the roller. In one embodiment, r=5. Further, It can be a feature vector after multi-scale convolution (such as probabilistic neural network).

[0076] 2) Determine the parameter constraints of the preset probabilistic neural network based on the preset roller mechanics equation.

[0077] 3) Determining initialization weight parameters of a preset probabilistic neural network based on the parameter constraint items and the main physical mode, so as to physically constrain the nonlinear elastic characteristics of the joint features based on the preset probabilistic neural network.

[0078] The roll mechanics equation is calculated based on the physical properties of the roll material and the strain hardening characteristics of the roll material. express. It is the discrete form of the physical equation of roll deformation, which is used to constrain the model parameters to conform to the nonlinear elastic characteristics of cold-formed steel, such as the quadratic term of the stress-strain curve. The calculation method is expressed as:

[0079]

[0080] in, The feature vector after multi-scale convolution is the feature output of the s-th sensor. The elastic coefficient of the first material is obtained based on the physical properties of the material and experimental data through static mechanical model or experimental fitting. For example, the elastic coefficient can be determined by stress-strain curve or finite element analysis and is set to 0.3; is the elastic coefficient of the second material, which is obtained based on the nonlinear stress response of the material or polynomial fitting, and depends on the strain hardening characteristics of the material. For example, it can be obtained by fitting the nonlinear behavior in the experimental data or the material mechanics model, and is set to 0.5.

[0081] Furthermore, parameter constraints are constructed based on prior knowledge of roll mechanics, and physical equations are embedded in the initialization process to initialize the weights of the probabilistic neural network. This ensures that the model's initial parameters conform to the physical laws of the roll forming process, accelerating convergence and improving generalization. Conventional probabilistic neural network weight initialization methods (such as Xavier initialization) are typically designed based solely on the statistical distribution of data and do not consider the physical constraints of specific application areas, which can easily cause the model's convergence direction to deviate from the actual process laws. In specific implementation, the calculation method for the initialization weight parameters of the embodiment of the present invention is expressed as:

[0082]

[0083] Where, For probabilistic neural network Initial values ​​of layer weights; is the physical constraint strength factor, preferably, set to 0.1. is the discrete form of the physical equation of roll deformation; Characterized by The moving average expectation of is the singular value decomposition projection of the feature distribution.

[0084] Step S218 , based on the mixed Gaussian distribution of the joint features, capturing the non-steady-state characteristics of the roll quality degradation process indicated by the joint features under physical constraints.

[0085] Step S220 : generating a quality category probability output result corresponding to the joint feature based on the non-steady-state characteristics.

[0086] Due to the limitations of sensor accuracy, noise interference, and environmental factors (such as temperature changes, mechanical vibrations, etc.) that easily affect the measurement results, the sensor data has uncertainty characteristics. During the cold-bent steel roll forming process, the uncertainty of sensor data may cause errors in the monitoring data, thereby affecting the accuracy of quality assessment. In order to model the uncertainty of sensor data, the traditional probabilistic neural network assumes that the output obeys a fixed distribution, which makes it difficult to capture the non-steady-state characteristics of the roll quality degradation process. In an embodiment of the present invention, after the nonlinear elastic characteristics of the joint features are physically constrained through the above-mentioned step S216, the non-steady-state characteristics in the joint features are captured, so that the quality of the corresponding roll in the forming process is evaluated based on the non-steady-state characteristics. Even if the quality of the cold-bent steel roll degrades during the forming process, the subtle changes and potential faults in the complex quality degradation process can be captured to perform timely and accurate quality assessment. In specific implementation, the non-steady-state characteristics are captured through the following steps:

[0087] 1) Determine the mixed Gaussian distribution components according to the typical degradation stages corresponding to the roller.

[0088] 2) Based on the time step mean and time step standard deviation corresponding to the mixed Gaussian distribution components, determine the normal distribution corresponding to the joint feature.

[0089] In one embodiment, typical degradation stages for cold-formed steel quality assessment include: normal, slight deformation, moderate wear, severe failure, and critical failure. Correspondingly, the number of components of the mixed Gaussian distribution is is 5 to indicate the number of Gaussian distributions used in the dynamic mixture density output layer. A Gaussian distribution (also known as a normal distribution) is a continuous probability distribution that peaks at its mean and decreases symmetrically on both sides.

[0090] 3) Based on the normal distribution, a preset probabilistic neural network is used to capture the non-steady-state characteristics of the roll quality degradation process indicated by the joint features.

[0091] In its implementation, the present invention utilizes a dynamic mixed density output layer for data processing. During this process, the non-steady-state characteristics corresponding to the joint features are determined to address the problem of sudden changes in quality degradation patterns. This is accomplished by obtaining the component contributions corresponding to the mixed Gaussian distribution components. Based on the component contributions and the final hidden state of a pre-set probabilistic neural network, the normal distributions are mixed to determine the probabilistic output of the quality category indicated by the joint features. Specifically, the corresponding evaluation and classification results are determined using the following formula:

[0092]

[0093] Where, Indicates that in a given (i.e. input features), the output probability. For the The model output at each time step represents the evaluation and classification results of the molding quality.

[0094] In an embodiment of the present invention, the output layer includes the mixing coefficients , the calculation method is expressed as ; is the Softmax classification function. This mixing coefficient is used to normalize the input of the dynamic mixed density output layer to avoid mode collapse caused by different scales. is the weight matrix of the mixing coefficient, which is a trainable parameter used to calculate the contribution of each component. is the last hidden state, representing the probabilistic neural network model in the first The time step is a high-dimensional representation of the current input, which contains a summary of all input information in the past moments.

[0095] Based on the above formula, the distribution parameters (i.e., mixing coefficients) of the output layer in this embodiment of the present invention are adaptively adjusted by the hidden state. In one embodiment, this can be achieved using an online EM algorithm (Expectation-Maximization). The EM algorithm iteratively updates the mixing coefficients (distribution parameters) using a sliding window, calculating the posterior probability in an E-step and updating the mean, variance, and weight in an M-step.

[0096] Represents output The mean is and the standard deviation is Normal distribution; is the mean value of the mth mixture component at the tth time step, and the calculation method is expressed as: . is the weight matrix of the mean, which is a trainable parameter. is the standard deviation of the mth mixture component at the tth time step, explicitly modeling the uncertainty of sensor data, such as vibration signal noise. The calculation method is expressed as: ; is the weight matrix of the standard deviation, which is a trainable parameter.

[0097] Step S222: Based on the quality category probability output result, the forming quality of the roller is evaluated.

[0098] In one embodiment, the embodiment of the present invention trains the above-mentioned probabilistic neural network by using a preset training sample set, thereby using the above-mentioned output layer of the probabilistic neural network to determine the corresponding quality category probability output result. The training sample set contains data labels, such as by specifically marking different roller states or quality levels, so as to determine the corresponding labels. In the embodiment of the present invention, the annotation of data is mainly based on the actual quality conditions during the cold-bent steel roller forming process. Correspondingly, the annotated categories include: normal state: the roller works within the expected quality range and no abnormalities occur; slight deformation: the roller has a slight deformation, which may be caused by material fatigue or other minor factors; moderate wear: the roller begins to show obvious wear or deformation, affecting the forming quality; serious fault: the roller has a serious mechanical fault or deformation, which may cause serious problems in the forming process; critical failure: the roller is about to fail completely, and an irreversible fault may occur. In one embodiment, the annotation method is manual annotation.

[0099] Another cold-bent steel roll forming quality assessment method provided by an embodiment of the present invention obtains corresponding forming process monitoring data after normalizing the data while taking into account the signal fluctuation characteristics and the coupling relationship between sensors. The embodiment of the present invention can eliminate sensor dimensional differences while explicitly modeling the physical correlation across sensors in the cold-bend forming process, avoiding feature fragmentation of multi-sensor data and enhancing the consistency of cross-sensor features. Furthermore, the embodiment of the present invention also designs a feature fusion efficiency comparison experiment to verify the effectiveness of dynamic spatiotemporal normalization on multi-sensor feature fusion. Figure 4 The figure shows the comparison results of the effectiveness of feature fusion corresponding to the normalization method of the embodiment of the present invention. Figure 4 As can be seen, the unnormalized data performs the worst, verifying the necessity of normalization. Traditional normalization methods achieve a fusion efficiency of approximately 65%, while the normalization method of the embodiment of the present invention achieves a fusion efficiency of approximately 89%. This significantly outperforms traditional normalization methods (such as independent normalization). This demonstrates that the spatiotemporal covariance matrix-based normalization method of the embodiment of the present invention can effectively capture dynamic correlations between sensors and achieve effective feature fusion.

[0100] Furthermore, the embodiment of the present invention also embeds prior knowledge of roller mechanics into the initialization process of the neural network to implement physical constraints. By analyzing the impact of physical constraint initialization on the spatial distribution of neural network parameters and comparing it with the traditional random initialization method, the parameter distribution generated by the embodiment of the present invention presents a clustered form that is highly consistent with the mechanical properties of the material. The projection direction of the parameter vector in the manifold space maintains good consistency with the constraint direction of the roller deformation physical equation. Correspondingly, Figure 5 Figure 2 shows a schematic diagram of parameter space distribution and aggregation morphology. Figure 5 It can be seen that the embodiments of the present invention can effectively guide the network parameters to converge in a direction that conforms to actual physical laws, avoiding the problems of slow convergence and easy falling into local optimality caused by blind exploration of parameter space in traditional methods, and significantly improving the stability and convergence efficiency of model training.

[0101] Furthermore, the embodiment of the present invention also uses a physically constrained adaptive momentum optimization strategy and combines process prior knowledge to perform a parameter update process using the gradient descent method of a probabilistic neural network, so as to further fully capture subtle changes and potential failures in the complex quality degradation process. The calculation formula for parameter update is expressed as follows:

[0102]

[0103] In the above formula, For the Batch Probabilistic Neural Network layer weights; For the The learning rate of the batch to control the update step size. For the Batch Probabilistic Neural Network The momentum term of the layer; For the Batch Probabilistic Neural Network The adaptive learning rate term of the layer to control the influence of the momentum term; is a numerical stability constant, preferably, Set to . is the physical constraint fusion coefficient, preferably, Set to ; Characterizing the physical equations Batch Probabilistic Neural Network Sensitivity of layer weights.

[0104] In combination with the above formula, the embodiment of the present invention also designs adaptive momentum optimization parameters and adaptive learning rates for the above weight parameters, so as to determine the corresponding weight parameters. In specific implementation, the embodiment of the present invention determines the feature distribution difference corresponding to the joint feature based on the kernel density estimation distribution of the joint feature corresponding to each iteration of the probabilistic neural network; based on the feature distribution difference, the learning rate of the probabilistic neural network is adaptively adjusted. Among them, the embodiment of the present invention approximates the distribution of the entire data set by smoothing the density around each sample point. Even if the roller is in a non-steady-state working state or the data has distribution drift, it can be monitored and responded in real time. However, traditional learning rate schedulers all assume that the input data is independent and identically distributed. That is, it is assumed that each sample is extracted from the same probability distribution, and the samples are independent of each other. Based on the traditional learning rate scheduler, once the roller data has distribution drift between batches, it will lead to poor stability of cross-batch training.

[0105] The present invention uses an adaptive learning rate setting method based on process similarity to adjust the current learning rate using the similarity of feature distributions across historical batches. When the distribution changes suddenly between batches, the learning rate is automatically reduced to prevent parameter oscillation, improve cross-batch training stability, and adapt to the non-steady-state conditions in cold-formed steel manufacturing. Specifically, the adaptive learning rate is calculated as follows:

[0106]

[0107] Where, is the learning rate of the b-th batch. is the basic learning rate, preferably, Set to 0.0001. is an exponential function with a natural constant as its base; is the drift sensitivity coefficient, preferably, Set to 0.2. The Jensen-Shannon divergence measures the difference in feature distributions between adjacent batches and is used to detect process drift between roll batches (such as sensor distribution offset caused by material replacement) and automatically reduce the learning rate to prevent model oscillation. is the kernel density estimation distribution of the b-th batch features; is the kernel density estimation distribution of the b-1th batch features.

[0108] Furthermore, the above-mentioned adaptive momentum optimization parameters can be calculated based on the layer weight gradient of the probabilistic neural network, wherein the calculation formula of the adaptive momentum optimization parameters is expressed as:

[0109]

[0110] in, is the b-1th batch of probabilistic neural network The momentum term of the layer; For probabilistic neural network Layer weight gradients; is the momentum decay coefficient, preferably, Set to 0.9. Furthermore, each layer of the probabilistic neural network of the embodiment of the present invention also includes an adaptive learning rate term, and the adaptive learning rate can also be calculated based on the layer weight gradient of the probabilistic neural network. The calculation method is expressed as:

[0111]

[0112] in, is the b-1th batch of probabilistic neural network Adaptive learning rate term for the layer; is the adaptive term attenuation coefficient, preferably, Set to 0.999, this value can decay slightly with batch training, ensuring that there will be no large fluctuations and small adaptive adjustments.

[0113] During the training process of cold-bent steel roll forming quality assessment, the gradient direction of the model is easily interfered by noise sensors. To solve this problem, the existing technology usually adopts the method of randomly discarding nodes to make the model more sparse, retain more important features, and filter noise features to solve the noise interference problem. However, this method easily destroys the integrity of the physical constraints of the sensor data. In response to this, the embodiment of the present invention adopts a physically perceived directional discarding strategy, which dynamically shields irrelevant gradients according to the contribution of the nodes to the physical equations during back propagation, so as to suppress the propagation of noise gradients while retaining key physical constraints. Among them, the embodiment of the present invention determines the layer weight gradient of the probabilistic neural network based on the current node of the probabilistic neural network and the loss function of the probabilistic neural network. , and calculate the momentum term of the probabilistic neural network , and then applied to the weight update process to determine the final node. Specifically, the embodiment of the present invention performs physical perception probability discarding on the nodes of the probabilistic neural network based on the sensitivity of the preset roller mechanics equation to the layer weight of the probabilistic neural network, thereby determining the above-mentioned current node of the probabilistic neural network. It can be expressed as:

[0114]

[0115] Where, For probabilistic neural network Layer weight gradients; is the loss function of the probabilistic neural network; The loss function of the probabilistic neural network is The gradient of the layer weights reflects the impact of the model parameters on the loss. It is the Hadamard product, that is, element-level multiplication, which means element-by-element operation; is the symbol of partial derivative; The probabilistic neural network Layer weight matrix.

[0116] is the Dropout discard strength of the probabilistic neural network, preferably, Set to 0.1. Characterizing the physical equations Batch Probabilistic Neural Network The sensitivity of the layer weights characterizes the protection of weights that are strongly related to the roller mechanics equations (such as the neuron connections corresponding to the elastic coefficients of the first material and the elastic coefficients of the second material), and suppresses the gradient disturbance of the noise sensor. It represents the maximum value of the input data, which is used to select the maximum value of the weight sensitivity to the physical constraint.

[0117] Furthermore, for the cold-bent steel roll forming quality assessment task, the loss function adopted by the embodiment of the present invention takes into account the simultaneous optimization of classification accuracy and distribution calibration ability. In the specific implementation, the present invention determines the distribution density penalty term corresponding to the joint feature based on the standard deviation of the mixed Gaussian distribution component; and determines the loss function of the probabilistic neural network based on the modified negative log-likelihood loss and the distribution density penalty term. The embodiment of the present invention determines the corresponding loss function by adopting the modified negative log-likelihood loss and the distribution density penalty term. Specifically, the calculation method of the loss function is expressed as:

[0118]

[0119] Where, is the loss function of the probabilistic neural network; is the total length of the time series; is the number of components of the mixed Gaussian distribution, indicating the number of Gaussian distributions used in the dynamic mixed density output layer. Preferably, Set to 5. The penalty term that characterizes the distribution density corresponding to the joint features means that by penalizing the weighted sum of the variance, the model is forced to maintain low uncertainty in the accurate area and explicitly express high uncertainty in the fuzzy area. To control the hyperparameter of the calibration strength, preferably, Set to 0.3.

[0120] In summary, the embodiments of the present invention perform model training using the aforementioned loss function, adaptive learning rate, weight gradient, and adaptive momentum optimization parameters. Model training is completed by repeatedly iterating the aforementioned steps until a preset stop iteration condition is met. For example, the stop iteration condition may be set to reach a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000.

[0121] Furthermore, an embodiment of the present invention also provides a cold-bent steel roll forming quality assessment device, Figure 6 The schematic diagram of the structure of a cold-bent steel roll forming quality assessment device provided by an embodiment of the present invention is shown. Figure 6 The device includes: a data monitoring module 100, which is used to monitor the time series data of the physical quantity of the preset rolling mill during the forming process to obtain the forming process monitoring data of the preset rolling mill; the forming process monitoring data includes time series data corresponding to multiple sensors; a feature extraction module 200, which is used to use a preset wavelet packet basis function to determine the typical frequency band characteristics between the time series data of the forming process monitoring data, and based on the typical frequency band characteristics, perform multi-scale convolution processing on the forming process monitoring data to determine the joint characteristics representing local mutations and global trends in the forming process monitoring data; a data processing module 300, which is used to perform physical constraint processing on the nonlinear elastic characteristics of the joint characteristics according to the preset rolling mill mechanics equation; and, based on the mixed Gaussian distribution of the joint characteristics, capture the non-steady-state characteristics of the rolling mill quality degradation process indicated by the joint characteristics under physical constraints; an execution module 400, which is used to generate a quality category probability output result corresponding to the joint characteristics based on the non-steady-state characteristics; and an output module 500, which is used to perform quality assessment on the forming quality of the rolling mill based on the quality category probability output result.

[0122] An embodiment of the present invention provides a cold-bent steel roll forming quality assessment device, the implementation principle and technical effects of which are the same as those of the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the device embodiment, reference may be made to the corresponding contents in the aforementioned method embodiment.

[0123] Furthermore, the feature extraction module 200 is also used to: cluster the historical data spectrum of the molding process monitoring data, take the center of the cluster as the center frequency of the wavelet packet basis function; and determine the typical frequency band characteristics corresponding to the molding process monitoring data based on the center frequency.

[0124] The above-mentioned data monitoring module 100 is also used to: use multiple sensors to collect data from preset key positions of the preset rolling mill during the forming process, and obtain the initial monitoring time series data of each sensor at the preset key position; maximize the mutual information between the initial monitoring time series data of each sensor, and determine the lag time step corresponding to the initial monitoring time series data; determine the time series lag correlation corresponding to the initial monitoring time series data based on the lag time step; based on the time series lag correlation and the time window sliding variance of the initial monitoring time series data, jointly normalize the initial monitoring time series data to determine the forming process monitoring data of the preset rolling mill.

[0125] The above-mentioned data processing module 300 is also used to: perform singular value decomposition projection on the time window sliding average expectation of the joint feature to extract the main physical mode of the joint feature; determine the parameter constraint terms of the preset probabilistic neural network based on the preset roller mechanics equation; wherein the roller mechanics equation is calculated based on the material physical properties and material strain hardening characteristics of the roller; determine the initialization weight parameters of the preset probabilistic neural network based on the parameter constraint terms and the main physical mode, so as to physically constrain the nonlinear elastic characteristics of the joint feature based on the preset probabilistic neural network.

[0126] The above-mentioned data processing module 300 is also used to: determine the mixed Gaussian distribution components based on the typical degradation stages corresponding to the rolling mill rolls; determine the normal distribution corresponding to the joint features based on the time step mean and time step standard deviation corresponding to the mixed Gaussian distribution components; based on the normal distribution, use a preset probabilistic neural network to capture the non-steady-state characteristics of the rolling mill roll quality degradation process indicated by the joint features; wherein the weight parameters of the probabilistic neural network are initialized based on the physical constraint terms corresponding to the rolling mill roll mechanics equations.

[0127] The above-mentioned execution module 400 is also used to: obtain the component contribution corresponding to the mixed Gaussian distribution component; mix the normal distribution according to the component contribution and the last layer hidden state of the preset probabilistic neural network to determine the quality category probability output result indicated by the joint feature.

[0128] Among them, the weight parameters of the preset probabilistic neural network are determined based on the preset adaptive momentum optimization parameters and adaptive learning rate; the adaptive momentum optimization parameters and / or adaptive learning rate are calculated based on the layer weight gradient of the probabilistic neural network; the above-mentioned execution module 400 is also used to: based on the preset roller mechanics equation to the layer weight sensitivity of the probabilistic neural network, perform physical perception probability discarding on the nodes of the probabilistic neural network; based on the current node of the probabilistic neural network and the loss function of the probabilistic neural network, determine the layer weight gradient of the probabilistic neural network.

[0129] The above-mentioned execution module 400 is also used to: determine the distribution density penalty term corresponding to the joint feature based on the standard deviation of the mixed Gaussian distribution component; and determine the loss function of the probabilistic neural network based on the modified negative log-likelihood loss and the distribution density penalty term.

[0130] The execution module 400 is further configured to: determine the feature distribution difference corresponding to the joint feature based on the kernel density estimation distribution of the joint feature corresponding to each iteration of the probabilistic neural network; and adaptively adjust the learning rate of the probabilistic neural network based on the feature distribution difference.

[0131] An embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned Figures 1 to 2 The embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to execute the above Figures 1 to 2 Any of the steps of the method shown.

[0132] The embodiment of the present invention also provides a structural diagram of an electronic device, such as Figure 7 FIG. 1 is a schematic diagram of the structure of the electronic device, wherein the electronic device includes a processor 71 and a memory 70, the memory 70 stores computer executable instructions that can be executed by the processor 71, and the processor 71 executes the computer executable instructions to implement the above Figures 1 to 2 Either of the methods shown. Figure 7 In the illustrated embodiment, the electronic device further includes a bus 72 and a communication interface 73 , wherein the processor 71 , the communication interface 73 and the memory 70 are connected via the bus 72 .

[0133] Among them, the memory 70 may include high-speed random access memory (RAM), and may also include non-volatile memory (non-volatile memory), such as at least one disk storage. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 73 (which can be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. can be used. The bus 72 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc., or an AMBA (Advanced Microcontroller Bus Architecture, on-chip bus standard) bus, wherein AMBA defines three types of buses, including APB (Advanced Peripheral Bus) bus, AHB (Advanced High-performance Bus) bus and AXI (Advanced eXtensible Interface) bus. The bus 72 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 7 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0134] The processor 71 may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by hardware integrated logic circuits in the processor 71 or by software instructions. The above processor 71 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of the present application can be directly implemented and executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium mature in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor 71 reads the information in the memory and combines its hardware to complete the above Figures 1 to 2 Any of the methods shown.

[0135] A computer program product for a cold-bent steel roll forming quality assessment method and apparatus provided in an embodiment of the present invention includes a computer-readable storage medium storing program code. The program code includes instructions that can be used to execute the methods described in the aforementioned method embodiments. For specific implementations, please refer to the method embodiments and will not be described in detail here. Those skilled in the art will clearly understand that, for ease and brevity of description, the specific operating processes of the system described above can refer to the corresponding processes in the aforementioned method embodiments and will not be described in detail here. Furthermore, in the description of the embodiments of the present invention, unless otherwise specified or limited, the terms "installed," "connected," and "connected" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; mechanical connections, electrical connections, direct connections, indirect connections through an intermediary, or internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances. If the functions described are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or the portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0136] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0137] Finally, it should be noted that the above embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. Such modifications, changes or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for evaluating the quality of cold-formed steel roll forming, characterized in that: The method comprises: Monitoring the time series data of the physical quantities of the preset roller during the forming process to obtain the forming process monitoring data of the preset roller; the forming process monitoring data includes the time series data corresponding to the multiple sensors; Using a preset wavelet packet basis function to determine typical frequency band features between the time series data of the molding process monitoring data, based on the typical frequency band features, multi-scale convolution processing is performed on the molding process monitoring data to determine the joint features characterizing local mutations and global trends in the molding process monitoring data; According to a preset roll mechanics equation, the nonlinear elastic characteristics of the joint feature are subjected to physical constraint processing; and based on a mixed Gaussian distribution of the joint feature, the non-steady-state characteristics of the roll quality degradation process indicated by the joint feature under the physical constraint are captured; Based on the non-steady-state characteristics, generating a quality category probability output result corresponding to the joint feature; Based on the quality category probability output result, a quality assessment is performed on the forming quality of the roll.

2. The method according to claim 1, characterized in that The step of determining typical frequency band features between the time series data of the molding process monitoring data using a preset wavelet packet basis function includes: Clustering the historical data frequency spectrum of the molding process monitoring data, and taking the center of the cluster as the central frequency of the wavelet packet basis function; Based on the center frequency, typical frequency band characteristics corresponding to the molding process monitoring data are determined.

3. The method according to claim 1, characterized in that The step of monitoring the physical quantity time series data of the preset roller during the forming process to obtain the forming process monitoring data of the preset roller includes: Using multiple sensors to collect data from preset key positions of the preset roller during the forming process, and obtain initial monitoring time series data of each sensor at the preset key position; Maximizing the mutual information between the initial monitoring time series data of each sensor, and determining the lag time step corresponding to the initial monitoring time series data; Determining the time series lag correlation corresponding to the initial monitoring time series data according to the lag time step; Based on the time series lag correlation and the time window sliding variance of the initial monitoring time series data, the initial monitoring time series data is jointly normalized to determine the forming process monitoring data of the preset rolling mill.

4. The method according to claim 1, wherein The step of physically constraining the nonlinear elastic characteristics of the joint feature according to a preset roller mechanics equation comprises: Performing singular value decomposition projection on the time window sliding average expectation of the joint feature to extract the main physical mode of the joint feature; Determining parameter constraints of a preset probabilistic neural network according to the preset roller mechanics equation; wherein the roller mechanics equation is calculated based on the material physical properties and material strain hardening characteristics of the roller; According to the parameter constraint item and the main physical mode, the initialization weight parameters of the preset probabilistic neural network are determined to physically constrain the nonlinear elastic characteristics of the joint feature based on the preset probabilistic neural network.

5. The method according to claim 1, wherein The step of capturing the non-steady-state characteristics of the roll quality degradation process indicated by the joint feature under the physical constraint based on the mixed Gaussian distribution corresponding to the joint feature includes: determining a mixed Gaussian distribution component according to a typical degradation stage corresponding to the roll; Determining a normal distribution corresponding to the joint feature based on a time step mean and a time step standard deviation corresponding to the mixed Gaussian distribution components; Based on the normal distribution, a preset probabilistic neural network is used to capture the non-steady-state characteristics of the roll quality degradation process indicated by the joint feature; wherein the weight parameters of the probabilistic neural network are initialized based on the physical constraint terms corresponding to the roll mechanics equation.

6. The method according to claim 5, characterized in that The step of generating a quality category probability output result corresponding to the joint feature according to the non-steady-state characteristic includes: Obtaining component contributions corresponding to the mixed Gaussian distribution components; The normal distribution is mixed according to the component contribution and the last hidden state of the probabilistic neural network to determine the quality category probability output result indicated by the joint feature.

7. The method according to claim 6, characterized in that The weight parameters of the probabilistic neural network are determined based on preset adaptive momentum optimization parameters and adaptive learning rates; the adaptive momentum optimization parameters and / or the adaptive learning rates are calculated based on the layer weight gradients of the probabilistic neural network; the method further includes: Based on the sensitivity of the preset roller mechanics equation to the layer weights of the probabilistic neural network, the nodes of the probabilistic neural network are discarded according to the physical perception probability; Based on a current node of the probabilistic neural network and a loss function of the probabilistic neural network, a layer weight gradient of the probabilistic neural network is determined.

8. The method according to claim 7, characterized in that The calculation method of the loss function includes: Determining a distribution density penalty term corresponding to the joint feature based on a standard deviation of the mixed Gaussian distribution component; A loss function of the probabilistic neural network is determined based on the modified negative log-likelihood loss and the distribution density penalty term.

9. The method according to claim 7, characterized in that The method for determining the adaptive learning rate includes: determining a feature distribution difference corresponding to the joint feature based on a kernel density estimation distribution of the joint feature corresponding to each iteration of the probabilistic neural network; Based on the feature distribution difference, the learning rate of the probabilistic neural network is adaptively adjusted.

10. A cold-bent steel roll forming quality assessment device, characterized in that: The device comprises: A data monitoring module is used to monitor the time series data of the physical quantities of the preset roller during the forming process to obtain the forming process monitoring data of the preset roller; the forming process monitoring data includes time series data corresponding to multiple sensors; a feature extraction module, configured to determine typical frequency band features between the time series data of the molding process monitoring data using a preset wavelet packet basis function, and perform multi-scale convolution processing on the molding process monitoring data based on the typical frequency band features to determine joint features representing local mutations and global trends in the molding process monitoring data; a data processing module configured to perform physical constraint processing on the nonlinear elastic characteristics of the joint feature according to a preset roll mechanics equation; and to capture, based on a mixed Gaussian distribution of the joint feature, the unsteady-state characteristics of the roll quality degradation process indicated by the joint feature under the physical constraint; An execution module, configured to generate a quality category probability output result corresponding to the joint feature based on the non-steady-state characteristic; An output module is used to perform quality assessment on the forming quality of the roll based on the quality category probability output result.

Images