Real-time industrial quality detection method and device based on dual-channel cooperative recognition and storage medium
By employing a dual-channel collaborative recognition method, utilizing a high-frequency self-attention recurrent unit and a dual-channel collaborative long short-term memory network, the problem of real-time detection of high-frequency process variables and low-frequency quality indicators in industrial processes was solved, achieving high-precision and low-latency industrial quality inspection.
Patent Information
- Application Number
- CN202611295550.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-25
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies struggle to achieve real-time detection of high-frequency process variables and low-frequency quality indicators in industrial processes, resulting in low detection accuracy and difficulty in meeting the requirements of real-time optimization control. Furthermore, pure data-driven models suffer from insufficient parameter oscillation and online adaptability under dynamic operating conditions.
A dual-channel collaborative recognition method is adopted, which extracts fast-scale dynamic features through high-frequency self-attention recurrent units, combines dual-channel collaborative long short-term memory networks, integrates high-frequency and low-frequency data features, and establishes a dynamic detection model for industrial quality indicators to achieve collaborative updating of linear mechanisms and nonlinear relationships.
By continuously and collaboratively updating linear and nonlinear parameters within the same model structure, local optima and parameter oscillations are reduced, online detection latency is decreased, and the real-time accuracy and stability of industrial quality inspection are improved.
Smart Images

Figure CN123070482A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent detection technology for industrial processes, and in particular to a real-time industrial quality detection method, equipment, and storage medium based on dual-channel collaborative identification. Background Technology
[0002] In process industries such as mineral processing, metallurgy, chemical engineering, energy, pharmaceuticals, and materials, product quality indicators are often crucial for production scheduling, operational optimization, and closed-loop control. Taking concentrate grade in mineral processing as an example, this indicator reflects the overall quality status of the production process. However, its true value often relies on manual sampling, offline testing, or complex analytical instruments, resulting in long detection cycles and significant feedback lags, making it difficult to meet the needs of real-time optimization control.
[0003] Industrial sites typically allow for the continuous collection of numerous high-frequency process variables, such as temperature, pressure, flow rate, current, vibration, liquid level, and equipment operating status. However, quality indicators, production statistics, and some operational indicators are sampled at lower frequencies. The discrepancies in sampling periods, response hysteresis, and complex dynamic coupling relationships between high-frequency process variables and low-frequency quality indicators make it difficult to directly employ downsampling, interpolation, or fixed-window splicing methods, which can easily lead to information loss and affect the real-time detection accuracy of industrial quality indicators.
[0004] Among existing real-time detection methods, Long Short-Term Memory (LSTM) networks, gated recurrent units (GRUs), convolutional neural networks (CNNs), and deep learning models based on attention mechanisms have been used for industrial time series modeling. However, purely data-driven models struggle to fully utilize the linear mechanisms of the system, and are prone to local optima, parameter oscillations, and insufficient online adaptability under dynamic operating conditions, raw material fluctuations, or equipment status changes. Some system identification methods can characterize linear relationships, but they are insufficient for independently describing unknown high-order nonlinear dynamics in complex industrial processes.
[0005] Therefore, there is a need for a real-time industrial quality inspection method that can simultaneously utilize the rapid change information of high-frequency process data, the long-term feedback information of low-frequency quality indicators, and the linear mechanism information of the system identification model. Summary of the Invention
[0006] To address the aforementioned technical problems, this invention provides a real-time industrial quality inspection method, equipment, and storage medium based on dual-channel collaborative recognition. This invention collects and organizes industrial dual-frequency time-series data to establish a dynamic detection model for industrial quality indicators. It extracts rapid-scale dynamic features of high-frequency process variables using a high-frequency self-attention recurrent unit; constructs a dual-channel collaborative long short-term memory network and embeds it into the system recognition channel; and fuses high-frequency dynamic features, low-frequency input features, and the output of the system recognition model to obtain real-time detection results for industrial quality indicators.
[0007] The technical means employed in this invention are as follows: A real-time industrial quality inspection method based on dual-channel collaborative recognition includes: acquiring and aligning high-frequency process variable data and low-frequency quality indicator data in the industrial process according to a low-frequency sampling period, and constructing a rolling sliding window input sample for real-time detection modeling; establishing a dynamic detection model for industrial quality indicators, which includes a system recognition model for characterizing linear mechanistic relationships and a nonlinear dynamic model for characterizing unknown high-order nonlinear relationships; inputting the high-frequency process variable data into a high-frequency self-attention recurrent unit, and extracting fast-scale dynamic features from the high-frequency process variable data through grouped two-dimensional convolution, self-attention calculation, and self-loop state transfer; embedding the output channel of the system recognition model into a dual-channel collaborative long short-term memory network unit, so that the linear mechanistic features and nonlinear dynamic features are continuously and collaboratively updated in the same network structure; generating real-time detection results of industrial quality indicators under future prediction step size based on the fused high-frequency dynamic features, low-frequency input features, and system recognition model output; and training, updating parameters, and deploying the dynamic detection model for industrial quality indicators at the edge to achieve real-time online detection of industrial quality indicators.
[0008] Furthermore, the high-frequency process variable data includes temperature, pressure, flow rate, current, vibration, and liquid level collected by sensors, PLCs, and DCS; the low-frequency quality index data includes industrial quality indicators and production quality data obtained by manual sampling, offline testing systems, and data recording systems.
[0009] Furthermore, the dynamic detection model for industrial quality indicators uses both high-frequency process variable data and low-frequency quality indicator data as model inputs, and outputs the industrial quality indicators under the future prediction step size. The system identification model is used to characterize the linear mechanism relationship between low-frequency input variables and industrial quality indicators, while the nonlinear dynamic model is used to characterize the unknown high-order nonlinear relationship under the combined effect of high-frequency process variables and low-frequency input variables.
[0010] Furthermore, the high-frequency self-attention loop unit divides the high-frequency time series within each low-frequency sampling period into multiple data matrices, performs grouped two-dimensional convolution processing on each, flattens and aggregates the convolution outputs to generate a query matrix, a key matrix, and a value matrix, obtains the high-frequency feature matrix through dot product attention calculation, and performs loop state transfer in conjunction with the self-loop state factor.
[0011] Furthermore, the dual-channel collaborative long short-term memory network unit includes a forget gate, an input gate, an output gate, a memory state unit, and a system identification channel; the system identification channel receives low-frequency time series as input and combines it with the state unit, so that the gradient acts on both the linear system identification parameters and the nonlinear dynamic parameters during forward and backward propagation; the dynamic detection model is deployed on one or more of the following: edge server, industrial computer, control server, industrial cloud server, local workstation, or embedded computing node.
[0012] The present invention also provides an apparatus including a memory, a processor, a communication interface, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a real-time industrial quality inspection method based on dual-channel collaborative identification.
[0013] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a real-time industrial quality inspection method based on dual-channel collaborative recognition.
[0014] Compared with the prior art, the present invention has the following advantages: This invention can continuously and collaboratively update the identification parameters and nonlinear dynamic parameters of a linear system within the same model structure, reducing local optima and parameter oscillations caused by alternating optimization; at the same time, it extracts fast-scale information from high-frequency process data through high-frequency self-attention recurrent units and reduces online detection latency through edge-side deployment. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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 creative effort.
[0016] Figure 1 This is a flowchart of the real-time industrial quality inspection method based on dual-channel collaborative recognition in this invention.
[0017] Figure 2 This is a structural diagram of the dual-channel collaborative long short-term memory unit in this invention.
[0018] Figure 3 This is a diagram of the multi-layer dual-channel collaborative long short-term memory network structure in this invention.
[0019] Figure 4 This is a comparison chart of the prediction results of the MLP model in the embodiments of the present invention.
[0020] Figure 5 This is a comparison chart of the prediction results of the GRU model in the embodiments of the present invention.
[0021] Figure 6 This is a comparison chart of the prediction results of the LSTM model in the embodiments of the present invention. Detailed Implementation
[0022] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0025] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values of the components and steps set forth in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following drawings denote similar items; therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings.
[0026] like Figure 1As shown, this invention provides a real-time industrial quality inspection method based on dual-channel collaborative recognition, comprising: acquiring and aligning high-frequency process variable data and low-frequency quality index data in the industrial process according to a low-frequency sampling cycle, and constructing a rolling sliding window input sample for real-time detection modeling; in a preferred embodiment of this invention, the high-frequency process variable data includes temperature, pressure, flow rate, current, vibration, and liquid level collected by sensors, PLCs, and DCS; the low-frequency quality index data includes industrial quality indicators and production quality data obtained by manual sampling, offline testing systems, and data recording systems.
[0027] In implementation, industrial quality indicators can be modeled as a dynamic control system consisting of three interacting levels: operational decision-making, operational control, and process control. Their relationship is expressed by the following formula:
[0028] in, Indicates industrial quality indicators. Indicates operational indicators. This represents production statistics. Indicates control input, Indicates control output. , and This indicates the corresponding unknown disturbance. for The derivative of for The derivative of for The derivative of . This indicates the operational decision-making system. Indicates the operation and control system. This refers to a process control system.
[0029] Forecasting step size for the future The dynamic system of industrial quality indicators is represented as follows:
[0030] in, Collectively referred to as process variables, F(·) is an unknown nonlinear model. , and These represent the model order for different variables. This indicates the time step of the current low-frequency sampling period.
[0031] In a low-frequency sampling period Within, it will be based on the sampling period High-frequency process variables collected The sequence is constructed as a high-frequency process variable matrix. :
[0032] in, The time step represents the high-frequency sampling period. express The vector of high-frequency process variables at a given time. This indicates the sampling period multiplier.
[0033] Let the low-frequency production variables and low-frequency statistical variables be denoted as follows: and , This represents the mathematical transpose symbol. High-frequency process variables and low-frequency quality indicators are aligned based on low-frequency sampling time labels, and a scrolling sliding window is constructed for input samples.
[0034] A dynamic detection model for industrial quality indicators is established. This model includes a system identification model to characterize linear mechanistic relationships and a nonlinear dynamic model to characterize unknown higher-order nonlinear relationships. In a preferred embodiment, the dynamic detection model uses both high-frequency process variable data and low-frequency quality indicator data as input, and outputs the industrial quality indicators at future prediction step sizes. The system identification model characterizes the linear mechanistic relationship between low-frequency input variables and industrial quality indicators, while the nonlinear dynamic model characterizes the unknown higher-order nonlinear relationships under the combined influence of high-frequency process variables and low-frequency input variables.
[0035] In the operational decision-making level involving human experience rules, the difference in quality indicators and the range of production statistics can be mapped to changes in operational indicators:
[0036] in, This indicates the lower limit of the deviation between the actual and target values of the grade. This represents the target value for industrial quality indicators. This indicates the upper limit of the deviation between the actual value and the target value of the grade. Indicates the first The lower limit of the value range for each production statistical indicator. Indicates the first One production statistical indicator, Indicates the first The upper limit of the value range for each production statistical indicator. Indicates the first Changes in each operational indicator The variable scaling factor represents the decision-making process. The index number indicating the production statistics indicator. This represents the variable scaling factor for decision-making.
[0037] Therefore, the operational decision-making process based on rule-based reasoning can be further written in the form of proportional integral:
[0038] in, Indicates the first Each operational indicator The time setting value.
[0039] Since the target quality index is usually constant during stable operation, the above process can be expressed as the sum of a linear model and an unknown nonlinear term:
[0040] in, This represents the shift operator polynomial, here , Represents the parameters of the linear model. It represents an unknown nonlinear term.
[0041] When an industrial process is in steady-state operation, using the actual values of operational indicators to replace the decision values, the dynamic system identification model for industrial quality indicators can be written as follows:
[0042]
[0043] in, Indicates the first The shift operator polynomial corresponding to each operation index , This represents the row vector of linear parameters corresponding to production statistics. , This represents the unknown higher-order nonlinear modeling error. Represents a vector of production statistics. Indicates the first The linear coefficient of each operational indicator at the current moment. Indicates the first The linear coefficient of each operational indicator at the current moment. These correspond to the linear coefficients of the first through fifth production statistical indicators, respectively. This represents an unknown nonlinear function. Represents a process variable vector.
[0044] The parameters of the linear model are identified using the least squares method or recursive least squares method, and the estimated parameters are obtained. The above equation can then be rewritten as:
[0045] in, Represents linear parameters The estimated value, Represents linear parameters The estimated value, combined error term It consists of both system identification error and unknown higher-order modeling error:
[0046] The estimated value of the quality index at the future prediction step size can be obtained by using the change in the quality index:
[0047] in, It represents the change in industrial quality indicators within a unit time step. Indicates the future Predicted values of industrial quality indicators under different step sizes.
[0048] Combining the linear system identification model with the nonlinear error term, we obtain:
[0049]
[0050]
[0051] in, This represents an unknown higher-order nonlinear error term. This represents the system's identification model.
[0052] Assuming that the model order for all variables is n, the unknown higher-order error terms can be simplified to a nonlinear model with fast-scale process data as input:
[0053] in, This represents a nonlinear error model.
[0054] Therefore, this invention identifies the model through a system. With nonlinear error model Collaborative learning enables real-time industrial quality monitoring of industrial quality indicators.
[0055] High-frequency process variable data is input into a high-frequency self-attention recurrent unit. Fast-scale dynamic features in the high-frequency process variable data are extracted through grouped two-dimensional convolution, self-attention calculation, and self-loop state transfer. In a preferred embodiment of the present invention, the high-frequency self-attention recurrent unit divides the high-frequency time series within each low-frequency sampling period into multiple data matrices, performs grouped two-dimensional convolution processing on each, flattens and aggregates the convolution outputs to generate a query matrix, a key matrix, and a value matrix, obtains the high-frequency feature matrix through dot product attention calculation, and performs recurrent state transfer in conjunction with the self-loop state factor.
[0056] The high-frequency self-attention recurrent unit divides the high-frequency time series V(t) within each low-frequency sampling period into... Data matrix , where j=1,…, And a filter kernel is used for grouped two-dimensional convolution and flattening aggregation:
[0057] in, This represents a two-dimensional convolution operation. Indicates the first A grouped convolutional filter kernel, This represents the operation of flattening the convolution output matrix into a one-dimensional vector. Indicates the first The output of the data matrix after convolution. Indicates the first The feature vectors obtained after convolution and flattening of the data set. Indicates the first The feature vectors obtained after convolution and flattening of the data set. This represents the convolutional aggregation matrix after all grouped flattened features are aggregated.
[0058] Convolutional aggregation matrix The query matrix is obtained after linear transformation. Key matrix Sum matrix
[0059]
[0060] Then, the high-frequency feature matrix is calculated using the dot product attention mechanism:
[0061] in, The dimension of the key vector. This represents the dot product attention calculation function. Represents the query matrix. Represents the key matrix, Represents a value matrix, This represents the normalized exponential function, used to transform attention scores into a probability distribution. This is used to scale the dot product result. A self-loop state factor is further introduced. The high-frequency dynamic feature matrix is obtained as follows:
[0062] in, This represents the high-frequency dynamic feature vector output by the high-frequency self-attention recurrent unit. This represents the activation function. Represents the Hadamard product. This represents the trainable weight matrix of the output layer. This represents the output matrix after self-attention computation. Represents the self-loop state factor Perform hyperbolic tangent transform. This represents the trainable bias parameters of the output layer.
[0063] The output channel of the system identification model is embedded into a dual-channel collaborative long short-term memory network unit, enabling continuous collaborative updates of linear mechanistic features and nonlinear dynamic features within the same network structure. Specifically, as a preferred embodiment of this invention, the dual-channel collaborative long short-term memory network unit includes a forget gate, an input gate, an output gate, a memory state unit, and a system identification channel. The system identification channel receives low-frequency time series as input and combines it with the state unit, allowing the gradient to simultaneously act on both linear system identification parameters and nonlinear dynamic parameters during forward and backward propagation. The dynamic detection model is deployed on one or more of the following: edge server, industrial computer, control server, industrial cloud server, local workstation, or embedded computing node.
[0064] Dual-channel collaborative long short-term memory unit structure, such as Figure 2 As shown. Forgotten Gate Input gate Output gate Candidate memory states and the process of updating memory state Represented as:
[0065] in, This represents the Sigmoid activation function. Indicates the input gate and the current input The connection weight matrix, Represents high-frequency dynamic feature vectors. This indicates the hidden state of the input gate compared to the previous time step. The connection weight matrix, This indicates the hidden state at the previous moment. This represents the bias parameters of the input gate. This represents the weight matrix connected to the current input in the forget gate. This represents the weight matrix in the forget gate that connects to the hidden state from the previous time step. This represents the bias parameter of the forget gate. This represents the weight matrix in the output gate that is connected to the current input. This represents the weight matrix in the output gate that connects to the hidden state from the previous time step. This represents the bias parameter of the output gate. This represents the weight matrix that connects to the current input in the candidate memory state. This represents the weight matrix connecting the candidate memory state to the hidden state at the previous time step. Bias parameters representing candidate memory states. It indicates the memory state at the previous moment.
[0066] The low-frequency channel input consists of low-frequency operating parameters and production statistics:
[0067]
[0068] in, This represents the low-frequency input feature vector at the current moment. This represents a low-frequency input feature sequence matrix composed of low-frequency input features from multiple time points. Indicates the first The low-frequency input feature vector at time t, Indicates the current moment.
[0069] The system identifies the hidden state output of the channel and the nonlinear dynamic channel output together to form the hidden state output of the dual-channel collaborative long short-term memory unit:
[0070] in, Indicates nonlinear dynamic channel output. This indicates the output of the linear system identification channel. and This indicates the output of a dual-channel collaborative long short-term memory unit. Because... Embedded within the network, the gradient can simultaneously act on the linear system identification parameters and nonlinear dynamic parameters during forward and backward propagation.
[0071] Multi-layer dual-channel collaborative long short-term memory network structure such as Figure 3 As shown, the first layer of the network receives high-frequency dynamic features and low-frequency input features from the output of the high-frequency self-attention recurrent unit; subsequent network layers receive the output of the previous layer and continuously introduce low-frequency input features to enhance the model's attention to low-frequency quality feedback and production statistics.
[0072] Based on the fused high-frequency dynamic features, low-frequency input features, and system identification model output, real-time detection results of industrial quality indicators at the future prediction step size are generated. During implementation, the high-frequency dynamic features, low-frequency input features, and system identification model output are fused, and the predicted values of industrial quality indicators at the future prediction step size are obtained through the prediction output layer. The prediction results can be used for human-machine interface display, quality trend forecasting, operating parameter optimization, production scheduling auxiliary decision-making, or reference information for near-end control loops.
[0073] The dynamic detection model for industrial quality indicators is trained, its parameters are updated, and it is deployed at the edge to achieve real-time online detection of industrial quality indicators.
[0074] During the training phase, historical industrial data is used to construct training and testing sets. Hyperparameters such as the number of training epochs, learning rate, batch size, window length, number of network layers, and number of hidden nodes are set. The model performance is evaluated using root mean square error, mean absolute error, and relative squared error, as expressed below:
[0075]
[0076]
[0077] in, This represents the root mean square error. To represent the square root operation, This represents the total number of samples in the test set or training set. This represents the summation operation. Indicates the sample index number. Indicates the first Predicted values of industrial quality indicators for a sample. Indicates the first The true values of industrial quality indicators for a sample. This represents the mean of the true values of industrial quality indicators.
[0078] After training, save the optimal model parameters and deploy the model to edge servers, industrial computers, control servers, industrial cloud servers, local workstations, or embedded computing nodes.
[0079] In the online application phase, data acquisition controllers, PLCs, DCSs, sensors, or online detection devices collect high-frequency process variables in real time, while manual sampling, offline testing systems, data recording systems, or online analyzers provide low-frequency quality indicator data. Data transmission between devices and systems can be accomplished via hardwired connections, industrial Ethernet, local area networks, TCP / IP, OPC, Modbus, or message queues. The above deployment methods are merely illustrative and do not constitute a limitation on the scope of protection of this invention.
[0080] The present invention also provides a device including a memory, a processor, a communication interface, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a real-time industrial quality inspection method based on dual-channel collaborative identification.
[0081] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a real-time industrial quality inspection method based on dual-channel collaborative recognition.
[0082] Example 1 In this embodiment, an industrial quality index prediction experiment was conducted using actual production data from a domestic ore dressing plant. The experimental data included 30 variables, including 11 low-frequency input variables, 18 high-frequency input variables, and 1 low-frequency output quality index. The low-frequency data was sampled at a 60-minute interval, and the high-frequency data was sampled once per minute. The dual-frequency industrial data were aligned according to the low-frequency sampling time label, resulting in 13,043 sets of low-frequency data and 782,580 sets of high-frequency data.
[0083] In model implementation, the high-frequency data matrix can be divided into... Four groups are used for convolutional feature extraction, with the convolutional filter kernel size set to w=3. The window length n is selected from the set {5,10,15,20}, the number of hidden nodes h is selected from the set {32,64,96,128}, and the number of network layers L is selected from the set {1,2,3,4}. In practice, n=5, h=32, and L=1 can be selected.
[0084] To verify the effectiveness of the method of the present invention, it can be compared with the multilayer perceptron model, the gated recurrent unit model, and the long short-term memory network model. Figure 4 This is a comparison chart of the prediction results from the MLP model. Figure 5 This is a comparison chart of the prediction results from the GRU model. Figure 6 This is a comparison chart of the prediction results from the LSTM model. The comparison results can be used to illustrate that the present invention has higher prediction stability and real-time industrial quality inspection accuracy under complex dynamic working conditions.
[0085] This embodiment provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the aforementioned real-time industrial quality inspection method based on dual-channel collaborative identification. The computer-readable storage medium can be a read-only memory, random access memory, flash memory, hard disk, solid-state drive, optical disk, magnetic disk, removable storage medium, or other non-transitory storage medium capable of storing program instructions.
[0086] In practical deployment, the device can communicate with sensors, PLCs, DCSs, online detection devices, real-time industrial databases, offline testing systems, online analyzers, or human-machine interfaces. It can transmit process variables, quality indicators, model parameters, and test results via hardwired connections, industrial Ethernet, local area networks, TCP / IP, OPC, Modbus, or message queues. The above deployment architecture is merely illustrative and does not constitute a limitation on the scope of protection of this invention.
[0087] In implementation, the device can perform functions such as data acquisition and organization, dynamic detection model construction, high-frequency self-attention feature extraction, dual-channel collaborative recognition modeling, prediction output, edge deployment and communication, and model training and updating through program instructions. These functions can be implemented in whole or in part through software, hardware, or a combination of both. When implemented in hardware, they can be achieved through one or more of the following: industrial computers, edge servers, control servers, industrial cloud servers, local workstations, embedded computing nodes, programmable logic devices, or dedicated computing circuits.
[0088] As another aspect of the present invention, this embodiment provides a real-time industrial quality inspection device based on dual-channel collaborative identification. The device includes a processor, a memory, a communication interface, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the aforementioned method embodiments.
[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A real-time industrial quality inspection method based on dual-channel collaborative recognition, characterized in that, include: According to the low-frequency sampling period, high-frequency process variable data and low-frequency quality index data in the industrial process are acquired and aligned to construct a rolling sliding window input sample for real-time detection modeling; A dynamic detection model for industrial quality indicators is established, which includes a system identification model for characterizing linear mechanistic relationships and a nonlinear dynamic model for characterizing unknown high-order nonlinear relationships. The high-frequency process variable data is input into a high-frequency self-attention recurrent unit, and fast-scale dynamic features in the high-frequency process variable data are extracted through grouped two-dimensional convolution, self-attention calculation and self-loop state transfer. The output channel of the system identification model is embedded into a dual-channel collaborative long short-term memory network unit, so that linear mechanism features and nonlinear dynamic features are continuously and collaboratively updated in the same network structure. Based on the fused high-frequency dynamic features, low-frequency input features, and system identification model output, real-time detection results of industrial quality indicators under the future prediction step size are generated. The dynamic detection model for industrial quality indicators is trained, its parameters are updated, and it is deployed at the edge to achieve real-time online detection of industrial quality indicators.
2. The real-time industrial quality inspection method based on dual-channel collaborative recognition according to claim 1, characterized in that, The high-frequency process variable data includes temperature, pressure, flow rate, current, vibration, and liquid level collected by sensors, PLCs, and DCS; the low-frequency quality index data includes industrial quality indicators and production quality data obtained by manual sampling, offline testing systems, and data recording systems.
3. The real-time industrial quality inspection method based on dual-channel collaborative recognition according to claim 1, characterized in that, The dynamic detection model for industrial quality indicators uses both high-frequency process variable data and low-frequency quality indicator data as model inputs, and outputs the industrial quality indicators under future prediction step sizes. The system identification model is used to characterize the linear mechanism relationship between low-frequency input variables and industrial quality indicators, while the nonlinear dynamic model is used to characterize the unknown high-order nonlinear relationship under the combined effect of high-frequency process variables and low-frequency input variables.
4. The real-time industrial quality inspection method based on dual-channel collaborative recognition according to claim 1, characterized in that, The high-frequency self-attention recurrent unit divides the high-frequency time series within each low-frequency sampling period into multiple data matrices, performs grouped two-dimensional convolution processing on each, flattens and aggregates the convolution outputs to generate a query matrix, a key matrix, and a value matrix, obtains the high-frequency feature matrix through dot product attention calculation, and performs cyclic state transfer in conjunction with the self-loop state factor.
5. The real-time industrial quality inspection method based on dual-channel collaborative recognition according to claim 1, characterized in that, The dual-channel collaborative long short-term memory network unit includes a forget gate, an input gate, an output gate, a memory state unit, and a system identification channel; the system identification channel receives low-frequency time series as input and combines it with the state unit, so that the gradient acts on both the linear system identification parameters and the nonlinear dynamic parameters during forward and backward propagation. The dynamic detection model is deployed on one or more of the following: edge server, industrial computer, control server, industrial cloud server, local workstation, or embedded computing node.
6. A device, characterized in that, The method includes a memory, a processor, a communication interface, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the real-time industrial quality inspection method based on dual-channel collaborative identification as described in any one of claims 1 to 5.
7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the real-time industrial quality inspection method based on dual-channel collaborative recognition as described in any one of claims 1 to 5.
