An intelligent control system and method for workshop equipment state based on internet of things

Abstract

The application discloses a kind of workshop equipment state intelligent control system and method based on Internet of Things, comprising: by deploying vibration, temperature, current, acoustic emission, infrared thermal image and environment and so on multi-source sensor network, real-time acquisition equipment state data, and transmission to cloud platform by 5G network, utilize edge computing equipment to carry out normalization preprocessing to data, obtain equipment state vector, based on the vector, with neural differential equation as framework to build intelligent state evolution model, obtain the best optimization parameter by metagradient learning, and then construct dynamic graph network to carry out equipment state space-time evolution modeling, capture the best mutation state data, finally, by introducing the deep space-time convolution network of dynamic convolution kernel morphing mechanism to the adaptive extraction of mutation feature, output equipment start-stop instruction and operating parameter adjustment and so on control strategy data, realize the precise, real-time closed-loop control to workshop equipment state.

Description

Technical Field

[0001] This invention relates to the field of program control technology, and in particular to an intelligent control system and method for the status of workshop equipment based on the Internet of Things. Background Technology

[0002] The in-depth application of Industrial Internet of Things (IIoT) technology provides a multi-source data foundation for intelligent control of workshop equipment status. As a core link in manufacturing, the operating status of equipment in the workshop directly affects production efficiency and safety. However, the workshop environment is complex, and the changes in equipment status are multi-source, dynamic, and strongly coupled. Multiple parameters such as vibration, temperature, and current affect each other in the spatiotemporal dimension, which can easily lead to equipment failure or even safety accidents. At present, existing technologies for workshop equipment status control have obvious hierarchical limitations and urgently need systematic improvement.

[0003] At the basic control level, automation systems based on programmable logic controllers (PLCs) and traditional sensors start and stop equipment through preset rules and have real-time response capabilities. However, they cannot capture gradual changes in status and early signs of faults. Their control strategies are rigid and difficult to adapt to dynamic changes in production load, resulting in high false alarm rates, passive maintenance, and a lack of predictive intervention capabilities.

[0004] At the data-driven level, the method of combining SCADA system with machine learning achieves fault diagnosis through historical data analysis. However, its model relies on discrete sampled data, which makes it difficult to depict the continuous evolution of equipment status. At the same time, traditional methods treat equipment as independent units, ignore the spatiotemporal correlation between equipment, and cannot effectively model the fault propagation effect, resulting in insufficient globality and accuracy of status assessment.

[0005] At the level of intelligent monitoring, deep learning solutions attempt to extract features through networks such as CNN and RNN. However, CNN is difficult to model the network topology of devices, RNN has long-range dependency problems, existing model structures are fixed, feature extraction mechanisms lack adaptability to state changes, and computational complexity is high, making it difficult to meet the low-latency control requirements of the edge side.

[0006] In summary, existing technologies have core shortcomings in terms of model discretization, lack of spatiotemporal correlation, feature rigidity, and real-time performance. There is an urgent need to build an integrated intelligent system that combines multi-source perception, continuous modeling, spatiotemporal analysis, and adaptive control to achieve more accurate, real-time, and adaptive closed-loop control. Summary of the Invention

[0007] In order to overcome the above-mentioned defects of the prior art and to achieve the above objectives, the present invention proposes the following technical solution: A workshop equipment status control system based on the Internet of Things, characterized in that it includes: S1: Deploy multi-source sensors to collect equipment status data in the workshop, and obtain equipment status vectors through normalization preprocessing operations; S2: Based on the device state vector, the optimal optimization parameters are obtained through the intelligent state evolution model, and a dynamic graph network is constructed based on the optimal optimization parameters to perform spatiotemporal evolution modeling of the device state and capture the optimal mutation state data in the device state data. S3: Based on the optimal mutation state data, a dynamic convolution kernel deformation mechanism is introduced through a deep spatiotemporal convolutional network to output device control strategy data; S4: Based on the equipment control strategy data, control the status of workshop equipment.

[0008] The process in S1 of deploying multi-source sensors to collect workshop equipment status data and obtaining equipment status vectors through normalization preprocessing is as follows: Vibration sensors, temperature sensors, and current sensors are used as multi-source sensors, which also include acoustic emission sensors and environmental sensors. The data acquisition frequency is set to once per second, and the raw data points are collected via a 5G network. Real-time transmission to the cloud platform; The collected raw data points are normalized using the following formula:

[0009] in, This refers to the minimum value of the data determined based on historical data or a preset range. For the maximum value of the data, This is the normalized device state vector; The normalization process is as follows: the original data is filtered and denoised using an edge computing device, and the extreme values ​​of each sensor parameter are dynamically calculated or preset, wherein the vibration frequency v is... =0Hz, =10kHz, temperature of =0℃, =100℃, for each data point Element-by-element calculations are performed according to the formula to obtain the normalized device state vector. ,Will As input data for intelligent state evolution models.

[0010] The process in S2 of obtaining the optimal optimization parameters based on the device state vector through the intelligent state evolution model is as follows: Using neural differential equations as the basic framework of the model, the device state vector is... The dynamic changes are modeled as a continuous process, and the model's differential equation is expressed as:

[0011] Wherein, the function The network is implemented using a multilayer perceptron parameterization method. The network structure contains three hidden layers with 128, 64, and 32 neurons respectively. Fully connected layers are used between them. The hidden layer activation function is ReLU, and the output layer uses a linear activation function. The model parameters... Includes weight matrix and bias vector ,in and These correspond to the input and output dimensions, respectively. The intelligent state evolution model is lightweighted through model pruning and quantization, with a pruning rate of 30% and quantization accuracy of FP16, ensuring inference latency of less than 50ms on edge computing devices. The model employs a selective update mechanism, updating when the device state change rate falls below a threshold. When the value is 0.05, skip the current calculation cycle; The process of obtaining the optimal optimization parameters ϑ through metagradient learning includes: using the Xavier initialization method to preset the initial parameters. Input initial time data Then, the fourth-order Runge-Kutta method was used for numerical solution, with a step size of 0.01 seconds and a local truncation error of [missing value]. Obtain the predicted value Through the loss function Calculate the prediction error, where Represents the number of time series samples; The adaptive moment estimation optimization algorithm was used for parameter optimization, with the learning rate set to 0.001 and the momentum parameter... =0.9, =0.999, numerical stability constant =10 -8 The iteration count is 1000, and the rate of change of the loss function is less than 10. -8 The optimization is stopped when the loss on the validation set fails to decrease after 10 consecutive iterations to prevent overfitting, thus obtaining the optimal optimization parameters. .

[0012] The process of building a dynamic graph network based on the optimal parameters in S2 to model the spatiotemporal evolution of device states is as follows: Workshop equipment is mapped as graph nodes, and physical connections or data flow associations between equipment are mapped as edges, defining edge weights. Reflecting the degree of influence between nodes, the weight calculation is based on the physical distance between devices and the frequency of data transmission; Based on optimal optimization parameters Dynamically update node attributes The updated formula is:

[0013] in, Indicates the first The set of neighbors of a node. The time step is set to 0.1 seconds; The specific process of spatiotemporal evolution modeling includes: establishing an adjacency matrix based on the device layout, recording node connection relationships to construct a graph structure, and calculating edge weights in real time. ,in The distance between devices. For data transmission frequency, through the spatiotemporal evolution equation Numerical solution is performed, where This is an identity function or linear transformation function used to process the state information of neighboring nodes and set a threshold for vibration frequency changes. =100Hz / s, temperature change threshold =5℃ / s, and when the threshold is exceeded, it is recorded as the optimal mutation state data.

[0014] The process in S3 of introducing a dynamic convolution kernel deformation mechanism based on optimal mutation state data through a deep spatiotemporal convolutional network is as follows: Constructing Deformation Networks The deformation network consists of two convolutional layers and a deformation parameter generator, with the following specific structure: First convolutional layer: kernel size is 3×3, number of channels is 64, stride is 1, padding is 1; The second convolutional layer has a kernel size of 5×5, 128 channels, a stride of 1, and padding of 2. Deformation parameter generator: consists of three fully connected layers with 256, 128, and 6 neurons respectively; The implementation process includes the following steps: processing the input optimal mutation state data through the first convolutional layer. Preliminary feature extraction is performed, among which In terms of time dimension, For spatial dimensions, For the channel dimension, features are further extracted through a second convolutional layer to obtain deep abrupt change state features. Based on deep mutation state characteristics The attention weights of sub-mutation features are calculated using an attention mechanism. The calculation formula is:

[0015] in The learnable parameter matrix; Based on device feature encoding vector Generate deformation parameters :

[0016] Through deformation parameters Adjusting the sampling coordinates of the convolution kernel enables elastic deformation of the convolution kernel, thus optimizing the feature extraction process; The optimal mutation state data is input in tensor form, with its dimensions matching the real-time data stream. Before feature extraction, the optimal mutation state data is concatenated with the real-time device state data of the current time window in the channel dimension as the joint input of the deep spatiotemporal convolutional network, enabling the model to simultaneously perceive instantaneous mutations and continuous states. Finally, the device control strategy data, including device start-stop commands and operating parameter adjustments, is output through a fully connected layer.

[0017] The control method comprises the following steps: S1: Deploy multi-source sensors to collect equipment status data in the workshop, and obtain equipment status vectors through normalization preprocessing operations; S2: Construct an intelligent state evolution model based on the device state vector to obtain the optimal optimization parameters; S3: Construct a dynamic graph network based on the optimal parameters to model the spatiotemporal evolution of equipment states and capture the data of the best abrupt change states; S4: Based on the optimal mutation state data, a dynamic convolution kernel deformation mechanism is introduced through a deep spatiotemporal convolutional network to output equipment control strategy data and realize the control of workshop equipment status.

[0018] The data acquisition module includes multi-source sensors such as vibration sensors, temperature sensors, and current sensors. The data is preprocessed by an edge computing device and then transmitted to the cloud platform.

[0019] The dynamic convolution kernel deformation mechanism in the control output module is implemented through a deformation network. The deformation network structure consists of two convolutional layers: the first convolutional kernel has a size of 3×3, and the second convolutional kernel has a size of 5×5. The deformation parameter generator adopts a fully connected layer structure.

[0020] The present invention has the following beneficial effects: In this invention, firstly, by deploying a multi-source sensor network including vibration, temperature, and current sensors, combined with 5G transmission and edge computing technologies, real-time acquisition and efficient preprocessing of equipment status data are achieved. A neural differential equation framework is used to accurately model the continuous dynamics of the equipment. Lightweight model processing reduces edge-side inference latency to the millisecond level, significantly improving the real-time performance of status monitoring. The introduction of a dynamic graph network effectively captures the spatiotemporal correlation characteristics between devices, achieving sub-second response speeds for detecting sudden changes in vibration frequency and temperature, and significantly improving spatial positioning accuracy. Secondly, the improved deep spatiotemporal convolutional network significantly enhances the flexibility of feature extraction through an adaptive deformation mechanism. This mechanism can dynamically adjust the receptive field according to changes in device state, and combined with a multi-level attention weighting strategy, it greatly improves the system's accuracy in identifying various abnormal operating conditions. Compared with traditional methods, the introduction of a dynamic graph network to model the correlation between devices is expected to reduce false alarms caused by noise from isolated devices. Combined with the adaptive deformation mechanism, it helps to enhance the ability to diagnose early faults, and the intelligent decision-making mechanism based on attention fusion can optimize control efficiency. Finally, an intelligent decision-making mechanism based on attention fusion optimizes the ratio of spatiotemporal characteristics to risk patterns. Through a dynamic weight allocation strategy, the system can autonomously identify key risk signals and accurately generate equipment start-up and shutdown commands and operating parameter adjustment schemes. This innovation significantly improves the overall control efficiency of the system, realizing a complete closed loop from state perception to decision execution, and providing strong support for the safe and stable operation of workshop equipment. Attached Figure Description

[0021] Figure 1 This is a system block diagram of an intelligent control system and method for workshop equipment status based on the Internet of Things proposed in this invention.

[0022] Figure 2 This is a flowchart illustrating the steps of an intelligent control system and method for workshop equipment status based on the Internet of Things proposed in this invention. Detailed Implementation

[0023] 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. 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] Example 1: As Figure 1 As shown, the present invention proposes an Internet of Things-based workshop equipment status control system, comprising: Data acquisition module: When deploying sensors, they need to be physically installed according to the equipment layout to ensure good contact between the sensors and the equipment surface or environment. Vibration sensors are usually installed on the equipment bearings or vibration-sensitive parts, temperature sensors are close to the equipment surface, current sensors are integrated into the power line, acoustic emission sensors are close to the bearing housing, infrared thermal imagers are aimed at the equipment surface, and environmental sensors are placed in the workshop environment. All sensors are connected to the data acquisition unit via wired or wireless means and are calibrated to ensure data accuracy.

[0025] After the sensors are deployed, the data acquisition module starts a continuous data acquisition process, with the acquisition frequency set to once per second, meaning each sensor acquires data once per second, forming raw data points. For the device state vector, the core data points include vibration frequency. ,temperature and current Therefore, the focus of data collection is on these three parameters. The original data point x is defined as a triple, i.e. The data collection process is synchronous, ensuring that the data collected is from the same timestamp. , and The value corresponds to the same device status.

[0026] Data acquisition is accomplished by an analog-to-digital converter inside the sensor, which converts analog signals into digital values, such as vibration frequency. Temperature is measured by a vibration sensor using the piezoelectric effect. The current is measured by a thermocouple or resistance temperature detector. Raw data points collected by Hall effect sensor. It is temporarily stored in the sensor's buffer, awaiting transmission.

[0027] The collected raw data points need to be transmitted to the processing platform via the network. The data transmission process first involves transferring the raw data points from the sensor buffer. The data is packaged into data packets, each containing a timestamp and sensor ID information. The data packets are then wirelessly transmitted to the cloud platform via a 5G modem. The transmission protocol uses a reliable TCP / IP protocol to ensure data integrity and order. The transmission frequency is consistent with the acquisition frequency, that is, one data packet is transmitted every second.

[0028] The advantages of 5G networks lie in their millisecond-level latency and gigabit-per-second bandwidth, making them suitable for large-volume data transmission in workshop environments. During transmission, data encryption technology is used to ensure security and prevent unauthorized access.

[0029] After receiving the transmitted data packets, the cloud platform performs data reception and storage processing. The reception process includes unpacking the data packets and extracting the raw data points. The system includes timestamps and verifies data integrity. If data packets are corrupted or lost, a retransmission mechanism ensures that no data is lost. Received data points are stored in the cloud platform's time-series database, which is designed to efficiently process time-series data and supports fast querying and retrieval.

[0030] During storage, the data is indexed by device ID and timestamp for easy subsequent processing. The result of this step is a sequence of raw data points stored in the cloud platform. These data points will be used for filtering and noise reduction. The stored data remains in its original form without any modification, ensuring that subsequent processing is based on accurate source data.

[0031] The stored raw data points may contain noise and outliers, thus requiring filtering and denoising. This processing is performed using edge computing devices located at the edge nodes of the cloud platform, close to the data source to reduce processing latency. The purpose of filtering and denoising is to remove high-frequency noise and transient interference, improving data quality. Specific filtering methods are based on digital filtering techniques and employ common practices: for each parameter... , and A one-dimensional moving average filter is applied.

[0032] The filtering process is performed on each data point. Let's proceed, assuming the data point at the current timestamp t is... The filter window size is s (e.g.) =5), then the filtered data points Calculated as past The average of a number of data points, mathematically, for a parameter Filtered value for:

[0033] Similarly, and The filter values ​​are calculated in the same way, and the filter window size is the same. Based on data volatility settings, typically =5 to balance smoothness and response speed. After filtering, random noise in the data is suppressed, and the trend is smoother.

[0034] To perform normalization, it is necessary to determine the minimum value of each parameter. and maximum value For vibration frequency The preset extreme value is =0Hz and =10kHz, for temperature The preset extreme value is =0℃ and =100℃, for current Because the power consumption of the device varies greatly and there is no preset range, dynamic calculation based on historical data is adopted.

[0035] The dynamic calculation process uses a filtered sequence of data points, and the extreme values ​​are determined individually for each parameter, such as current. Extreme values ​​are calculated based on data from a past period, such as a 24-hour data window. and Calculate them separately for the window. The minimum and maximum values, mathematically, are defined by the time window as... s, then:

[0036]

[0037] Window size Based on the equipment's operating cycle, it is usually set. =86400s (24h) to cover daily variations, for and By directly using the preset extreme values, no calculation is required. The result of this step is the extreme value pair for each parameter: , , These extreme values ​​will be used in the normalization calculation steps.

[0038] Normalization transforms the filtered data points into standardized device state vectors, scaling all parameter values ​​to the [0,1] range to eliminate the influence of dimensions. The normalization formula, based on a standard formula, is applied to each filtered data point. For each parameter, the normalized value is calculated as follows:

[0039]

[0040]

[0041] in, , and These represent the normalized vibration frequency, temperature, and current values, respectively. The normalization calculation is performed element-wise, meaning each parameter is processed independently. After calculation, if any value exceeds the range [0,1], it is truncated to the nearest boundary. However, since extreme values ​​are based on the data range, this situation is rare.

[0042] Finally, the normalized device state vector As the output of the data acquisition module, it is passed to the subsequent model building module. The output process includes... The data is packaged with a timestamp and device ID and stored in the cloud platform's output database. The output frequency matches the acquisition frequency, outputting a vector every second, which is the device status vector. It serves as the input for intelligent state evolution models and forms the basis for spatiotemporal evolution modeling.

[0043] Model building module: The first step of the model building module is to receive the device state vector from the data acquisition module. The device state vector is denoted as... Where t represents the timestamp, and the vector contains three normalized parameters: vibration frequency. ,temperature and current ,Right now The receiving process is completed through the cloud platform's data interface. The module reads the latest or historical time series data from the time series database. The verification process ensures the integrity and rationality of the vector data, such as checking whether the values ​​are in the range of [0,1] (due to normalization), whether the timestamps are continuous, and whether there are missing values. If an anomaly is found, the module will request the data acquisition module to resend or perform interpolation processing. The result of this step is a verified device state vector sequence, which will be used as the input data for the intelligent state evolution model. Verification is crucial because the model's performance depends on high-quality input, and all subsequent steps are based on this vector sequence.

[0044] After receiving the device state vector, the model building module initializes the basic framework of the intelligent state evolution model, namely the neural differential equation. The neural differential equation models the dynamic changes of the device state vector as a continuous process. Its core is a differential equation, which is expressed as:

[0045] Among them, the function Implemented by parameterization of multilayer perceptron. These are the model parameters, including the weight matrix and bias vector. However, the weight-related formulas are not expanded as required. The initialization process includes defining the form of the differential equation and setting the structure of the multilayer perceptron. The multilayer perceptron contains three hidden layers with 128, 64, and 32 neurons respectively. Fully connected layers are used between them. The hidden layer activation function is ReLU, and the output layer uses a linear activation function. The parameters... The initial value is preset by the Xavier initialization method, denoted as . Xavier initialization adjusts the weights based on the input and output dimensions to maintain gradient stability. The specific formula does not involve the details of the weights, but only describes it as random initialization.

[0046] Based on the initialized neural differential equation framework, the model building module performs numerical solutions to predict the evolution of the device state vector. The numerical solution uses the fourth-order Runge-Kutta method, and the solution process is performed at each time step with a step size of 0.01s to ensure computational accuracy. The input is the device state vector at the initial time step. ,in This is the start time of the sequence. The Runge-Kutta method calculates the next state value using multiple intermediate points, with a local truncation error of [missing information]. Where h is the step size, the specific calculation process is as follows: for each time step t, the predicted value is... Obtained through iterative calculation, using the current parameters. sum function The solution process covers the entire time series and generates a predicted sequence. For all time points, the result of this step is a predicted sequence of device state vectors. This sequence depends on the initial parameters and differential equation framework of the previous step and will be used in the loss calculation step.

[0047] After obtaining the predicted sequence, the model building module calculates a loss function to evaluate the error between the predicted and true values. The loss function is based on the mean squared error, and the formula is as follows:

[0048] in, Represents the number of time series samples. It is the actual device state vector. It's a predicted value, the loss. The average of the squared errors at all time points is calculated, reflecting the overall accuracy of the model's predictions. Error assessment also includes checking whether the loss value has converged, for example, by monitoring the rate of change of the loss.

[0049] Based on the loss value, the model building module uses an adaptive moment estimation optimization algorithm to optimize the parameters in order to minimize the loss function. The optimization process iteratively updates the parameters. The learning rate is set to 0.001, and the momentum parameter is... =0.9 and =0.999, numerical stability constant The iteration count is 1000, but an early stopping mechanism is used to prevent overfitting. The optimization process is as follows: In each iteration, the gradient is calculated, the first and second moment estimates are updated, and the parameters are adjusted. The specific formulas do not involve weights; only the process is described. The parameter update is based on a variant of gradient descent, using a loss function. The gradient information, when the rate of change of the loss function is less than Stop optimization when the time is right, or stop early if the loss on the validation set does not decrease after 10 consecutive iterations.

[0050] During the optimization process, the model building module simultaneously implements lightweight processing and selective update mechanisms to improve efficiency. Lightweight processing includes model pruning and quantization, with a pruning rate of 30% and quantization accuracy of FP16, ensuring inference latency of less than 50 milliseconds on edge computing devices. The selective update mechanism is based on the device state change rate: when the change rate is below a threshold... When the value is 0.05, the current calculation cycle is skipped, and the rate of change is calculated as the Euclidean distance between the device state vectors of adjacent time steps divided by the time interval.

[0051] Finally, the model building module outputs the optimal optimization parameters. Before outputting, a final verification is performed to ensure that the parameters perform well on the test set and meet the latency requirements. The optimal parameters are then used for the construction of the subsequent dynamic graph network.

[0052] Dynamic Network Module: The first step of the dynamic network module is to receive data input from the preceding modules. Specifically, the module receives the optimal optimization parameters from the model building module, denoted as... This parameter is obtained through optimization using an intelligent state evolution model and includes information such as the weight matrix and bias vector. However, details related to the weights are not elaborated upon as required. Simultaneously, the module receives the device state vector from the data acquisition module, denoted as... Where t represents the timestamp, and the vector contains the normalized vibration frequencies. ,temperature and current ,Right now The receiving process is completed through the cloud platform's data bus. The module accesses the shared storage area to read parameter and vector data. The verification process ensures the integrity and timeliness of the data, such as checking the optimal optimization parameters. Whether it has converged and is not empty, device state vector Whether the data is continuous over time and has a reasonable range (due to normalization, the values ​​should be between 0 and 1). If missing or anomalies are detected, the module will trigger a retransmission mechanism or interpolate using the most recent valid data. The result of this step is the verified optimal optimization parameters. and device state vector sequence These data will serve as the foundation for building the dynamic graph network. Validation is crucial because all subsequent steps depend on the accuracy of these input data, and any errors will lead to modeling bias.

[0053] After receiving and verifying the input data, the dynamic network module begins to construct the core graph structure of the dynamic graph network. The graph structure maps the physical equipment in the workshop to nodes in graph theory, and the physical connections or data flow associations between devices are mapped to edges. Specifically, each piece of equipment in the workshop (such as a machine tool, robot, or conveyor belt) is assigned a unique node identifier. The node attributes are initialized with the basic information of the equipment, such as equipment type, location coordinates, and operating status. Edges represent the associations between devices, such as devices connected by cables or pairs of devices that have data communication. The construction process is based on the workshop equipment layout diagram, which is pre-stored in the system database and contains the physical distances and connection relationships between devices. The module reads the layout diagram and initializes the adjacency matrix to record the node connection relationships. The adjacency matrix is ​​a square matrix, with rows and columns corresponding to node indices. If there is an edge between two nodes, the corresponding element of the matrix is ​​1; otherwise, it is 0. The directionality of the edges is determined according to the data flow direction, and an undirected graph is used by default.

[0054] Based on the constructed graph structure, the dynamic network module performs real-time calculation of edge weights, denoted as . ,in Indicates the edge index. The weight value represents the time interval and reflects the degree of influence between nodes. The calculation is based on the physical distance between devices and the data transmission frequency. The specific formula for weight calculation is as follows:

[0055] in, This refers to the distance between equipment units, in meters, obtained from the equipment layout diagram. This refers to the data transmission frequency between devices, measured in Hz, which is obtained in real time from the network monitoring system. This is the maximum transmission frequency in the system, a preset constant, for example, 1000Hz for normalization. The calculation process is performed in real time for each edge: the module queries the current timestamp. Value (by sampling network traffic data) and combined with static distance The weights are calculated by substituting them into the formula. The weight values ​​change dynamically, reflecting the real-time situation of network load and interactions between devices. After calculation, the weight values ​​are updated in the edge attributes of the graph structure. The result of this step is a weighted dynamic graph, where each edge has a corresponding weight. This weight will be used for node attribute updates. The weight calculation introduces a time dimension, enabling the graph network to capture dynamic changes.

[0056] After obtaining the weighted dynamic graph, the dynamic network module uses the optimal optimization parameters. Dynamically update node attributes, denoted as This indicates that the device is in time. State characteristics, initial attributes from device state vector Derivation, such as direct use As initial The update process is based on the update formula and utilizes the optimal optimization parameters. The formula for including neighbor node information is as follows:

[0057] in, This is the time step, set to 0.1s. Indicates the first The set of neighbors of each node, derived from the adjacency matrix. It is the set of weights of neighboring edges. This is the current device state vector. During updates, the module iterates through each node to obtain the current attributes. The neighbor weights and the current vector are then processed by a function. Calculate new attributes ,function The parameterization by the optimal optimization parameters is essentially a neural network model update where node attributes reflect the spatiotemporal evolution of the device state, incorporating the influence of local neighbors.

[0058] Based on the updated node attributes, the dynamic network module performs numerical solutions for spatiotemporal evolution modeling. The modeling uses a spatiotemporal evolution equation, which describes the changes in device state over time and incorporates the influence of the graph network. The equation is applied to each node. The definition is as follows:

[0059] Among them, the left side Represents a node Rate of change of state, first term on the right It is a local state evolution function that depends on the device state vector. Time t and optimal optimization parameters The second term is the neighbor influence term, summed. Neighbor nodes Weight and state vector Multiplication; numerical solutions are obtained using either the Euler method or the Runge-Kutta method. The simple Euler method is used to maintain real-time performance. The solution process involves calculating the rate of change for each time step. Then update the status. The solution covers all nodes and the entire time series, generating a predicted state evolution. The result of this step is the solved device state prediction value, which is used for mutation detection. The numerical solution discretizes the continuous equation to achieve computational feasibility.

[0060] After obtaining the predicted state values, the dynamic network module detects the optimal abrupt change state data. The abrupt change detection is based on the rate of change of vibration frequency and the rate of change of temperature, with the following thresholds set: Vibration frequency change threshold. Temperature change threshold =5℃ / s, the detection process is performed on each equipment node: the vibration frequency is extracted from the state prediction value. and temperature For the time series, the rate of change is calculated as an approximation of the derivative, i.e. and ,in =0.1s. If the rate of change exceeds the corresponding threshold, the state data recorded at the current timestamp is the best mutation state data. The mutation data includes the device identifier, timestamp, state vector and rate of change value. The detection is performed in real time. The module monitors streaming data and triggers the recording event.

[0061] Finally, the dynamic network module outputs the optimal mutation state data to the subsequent control output module. The output process includes data packaging and transmission: the mutation data is encapsulated into a standard format, containing metadata such as device ID and timestamp, and then sent to the control output module through a message queue. The output ensures low latency to meet real-time control requirements.

[0062] Control Output Module: The first step of the control output module is to receive optimal mutation state data from the dynamic network module. Optimal mutation state data consists of mutation state points detected by the dynamic network module during spatiotemporal evolution modeling. These points represent abnormal or critical changes in the device state, such as sudden changes in vibration frequency or temperature. The data is denoted as... It is a multidimensional tensor with dimension . ,in Represents the time dimension, indicating the length of the time series. Represents a spatial dimension, indicating the number of spatially distributed points of the device or sensor. The channel dimension represents the number of feature channels for each data point, such as vibration, temperature, and current. The receiving process is completed through the system's internal data pipeline, and the module reads data from the shared storage area. The data validation process ensures the integrity and reasonableness of the data, such as checking whether the tensor dimensions are consistent with expectations. Is it greater than zero? and Whether it matches the device layout and whether the value range is reasonable. Because mutated data may exceed the normal range, if a dimension error or missing data is found, the module will request the dynamic network module to resend or repair the data.

[0063] Receive and verify Subsequently, the control output module applies the first convolutional layer of the deep spatiotemporal convolutional network for preliminary feature extraction. The structure of the first convolutional layer is as follows: kernel size is 3x3, number of channels is 64, stride is 1, and padding is 1. Convolution is a linear transformation combined with a nonlinear activation function, but according to the requirements, the weight formula is not involved, so only the process is described. Specifically, the convolutional layer processes the input... A sliding window computation is performed, with each window having an inner product operation with the convolution kernel and then activated by the ReLU function, outputting a feature map F1. The dimension of the feature map F1 depends on... The dimensions and convolution parameters, for example if The size is take take After convolution, the size of F1 is take Multiply by 64, because the number of channels becomes 64, the convolution operation captures... The local spatiotemporal features, such as the temporal and spatial correlation of vibration modes, are calculated in parallel by GPUs or dedicated processors to ensure real-time performance.

[0064] Based on the feature map F1, the control output module applies a second convolutional layer to further extract features. The structure of the second convolutional layer is as follows: kernel size is 5x5, number of channels is 128, stride is 1, padding is 2. The convolutional layer performs a similar operation on F1 and outputs depth-abrupt state features. Because the convolution kernel is larger, the second layer captures a wider range of spatiotemporal context information, and the output... The dimension is take Multiply by 128, the number of channels increases to 128. It represents the deep features of the device status, integrates mutation information from multiple sources, and optimizes memory usage during the calculation process to avoid overflow.

[0065] In obtaining Then, the control output module calculates attention weights to emphasize... The key sub-features in the data, the attention mechanism is: using a learnable parameter matrix to... A linear transformation is performed, and then the weights are calculated using the softmax function. Specifically, the attention weights z1, z2, and z3 correspond to three sub-feature dimensions, such as vibration-dominated, temperature-dominated, and current-dominated abrupt change features. The calculation formula uses the standard softmax function, as shown below:

[0066] in yes The linear transformation result, The values ​​are 1, 2, and 3. Linear transformation involves a weight matrix. During calculation, the module first... Flatten it into a vector or preserve tensor form, then calculate. Finally, softmax is applied to obtain... Weight The values ​​are between 0 and 1, and the sum is 1, indicating the importance of each sub-feature.

[0067] Based on the attention weight vectors z1, z2, and z3, the output module is controlled to generate deformation parameters. The deformation parameter generator is a multilayer perceptron with a structure consisting of three fully connected layers with 256, 128, and 6 neurons respectively. The multilayer perceptron performs nonlinear transformations on the inputs z1, z2, and z3, and outputs... It is a 6-dimensional vector representing the offset parameters of the convolution kernel deformation. The computation process of the multilayer perceptron includes linear transformations and activation functions such as ReLU, but according to the requirements, it does not involve the weight formula. The weights have been learned during generator training, so the forward computation only requires matrix multiplication. The output is... Used to adjust the sampling coordinates of the convolution kernel.

[0068] Obtain deformation parameters Then, the control output module realizes the elastic deformation of the convolution kernel. The deformation process is specifically achieved by adjusting the coordinates of the sampling points in the standard convolution kernel grid. Let the coordinates of a sampling point in the standard grid be... New coordinates after deformation The calculation is as follows: ,in The offset is analytically derived from the deformation parameter γ. For non-integer coordinates, bilinear interpolation is used to obtain feature values, thereby achieving elastic deformation of the convolution kernel. This allows it to adapt to geometric changes in abrupt features. The deformation process adjusts the sampling coordinates of the convolution kernel, enabling it to adapt to irregular changes in the data. Specifically, for a standard convolution kernel, each sampling point has fixed coordinates; after deformation, the coordinates become the original coordinates plus... The provided offset, the amount of which is determined by The six elements control the possible spatial and temporal offsets. The deformation formula can be expressed as the new coordinates equal to the original coordinates plus the offset. After deformation, the convolution kernel is more flexible and can capture subtle features in abrupt changes.

[0069] Using deformed convolutional kernels, the output module is controlled to optimize feature extraction. The extraction process applies deformed convolutional kernels to features from a deep abrupt change state. Above this layer, additional convolution operations are performed, with convolution parameters similar to those of the previous layers, to output optimized features. , By capturing the features after deformation, the abrupt change state can be represented more accurately. During computation, the deformation convolution kernel and... Perform convolution operations and output a tensor. .

[0070] Based on optimization features The control output module applies the fully connected layer to output device control strategy data, and the fully connected layer will... Flattened into vectors, the control strategy is then output through linear transformations and activation functions such as sigmoid or tanh. The control strategy data includes equipment start / stop commands and operating parameter adjustments such as speed settings and temperature thresholds. The output dimension is set according to control requirements, such as binary commands or continuous values. The calculation of the fully connected layer involves a weight matrix, but the formula is not written as required. The result of this step is the equipment control strategy data, which will be sent to the actuator. The fully connected layer is the final layer of the network, which maps high-level features to control actions.

[0071] Finally, the control output module outputs control strategy data to the workshop equipment actuators. The output process includes data format conversion and transmission to ensure low latency. The actuators adjust the equipment status according to the strategy to complete the control cycle.

[0072] Example 2: Figure 2 As shown, the present invention proposes an intelligent control method for workshop equipment status based on the Internet of Things, comprising the following steps: S1: Multi-source data acquisition and preprocessing A multi-source sensor network, including vibration sensors, temperature sensors, and current sensors, is deployed on key equipment in the workshop. Sensor data is transmitted in real time to edge computing nodes via a 5G network for data cleaning, outlier processing, and feature extraction. A sliding window mechanism is adopted, with the window size set to 10 seconds and the overlap rate to 50%, to ensure the continuity and real-time nature of the data.

[0073] S2: Intelligent State Evolution Modeling The neural differential equation model is trained based on historical data to establish the temporal evolution law of device state. Through meta-learning strategy, the model can quickly adapt to the operating characteristics of different devices. Knowledge distillation technology is used to transfer the knowledge of complex models to lightweight models to ensure real-time reasoning capabilities on edge devices.

[0074] S3: Dynamic Graph Network Construction and Spatiotemporal Evolution Analysis Based on the workshop equipment layout and process flow diagram, a dynamic graph network model is constructed to analyze the spatiotemporal propagation characteristics of equipment status in real time. The graph attention network is used to capture the mutual influence relationships between equipment, and a temporal graph convolutional network is used to extract spatiotemporal features and identify abnormal propagation paths of equipment status.

[0075] S4: Intelligent Control Strategy Generation and Execution The optimal control strategy is generated based on a deep spatiotemporal convolutional network. The control effect is optimized through a model predictive control framework. The control parameters are adjusted online using a reinforcement learning algorithm to achieve adaptive control. The control strategy is then sent to the equipment control system via the OPCUA protocol to achieve closed-loop control.

[0076] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A workshop equipment status control system based on the Internet of Things, characterized in that, include: S1: Deploy multi-source sensors to collect equipment status data in the workshop, and obtain equipment status vectors through normalization preprocessing operations; S2: Based on the device state vector, the optimal optimization parameters are obtained through the intelligent state evolution model, and a dynamic graph network is constructed based on the optimal optimization parameters to perform spatiotemporal evolution modeling of the device state and capture the optimal mutation state data in the device state data. S3: Based on the optimal mutation state data, a dynamic convolution kernel deformation mechanism is introduced through a deep spatiotemporal convolutional network to output device control strategy data; S4: Based on the equipment control strategy data, control the status of workshop equipment.

2. The workshop equipment status control system based on the Internet of Things according to claim 1, characterized in that, The process in S1 of deploying multi-source sensors to collect workshop equipment status data and obtaining equipment status vectors through normalization preprocessing is as follows: Vibration sensors, temperature sensors, and current sensors are used as multi-source sensors, which also include acoustic emission sensors and environmental sensors. The data acquisition frequency is set to once per second, and the raw data points are collected via a 5G network. Real-time transmission to the cloud platform; The collected raw data points are normalized using the following formula: ； in, This refers to the minimum value of the data determined based on historical data or a preset range. For the maximum value of the data, This is the normalized device state vector; The normalization process is as follows: the original data is filtered and denoised using an edge computing device, and the extreme values ​​of each sensor parameter are dynamically calculated or preset, wherein the vibration frequency v is... =0Hz, =10kHz, temperature of =0℃, =100℃, for each data point Element-by-element calculations are performed according to the formula to obtain the normalized device state vector. ,Will As input data for intelligent state evolution models.

3. The workshop equipment status control system based on the Internet of Things according to claim 2, characterized in that, The process in S2 of obtaining the optimal optimization parameters based on the device state vector through the intelligent state evolution model is as follows: Using neural differential equations as the basic framework of the model, the device state vector is... The dynamic changes are modeled as a continuous process, and the model's differential equation is expressed as: ； Wherein, the function The network is implemented using a multilayer perceptron parameterization method. The network structure contains three hidden layers with 128, 64, and 32 neurons respectively. Fully connected layers are used between them. The hidden layer activation function is ReLU, and the output layer uses a linear activation function. The model parameters... Includes weight matrix and bias vector ,in and These correspond to the input and output dimensions, respectively. The intelligent state evolution model is lightweighted through model pruning and quantization, with a pruning rate of 30% and quantization accuracy of FP16, ensuring inference latency of less than 50ms on edge computing devices. The model employs a selective update mechanism, updating when the device state change rate falls below a threshold. When the value is 0.05, skip the current calculation cycle; The process of obtaining the optimal optimization parameters ϑ through metagradient learning includes: using the Xavier initialization method to preset the initial parameters. Input initial time data Then, the fourth-order Runge-Kutta method was used for numerical solution, with a step size of 0.01 seconds and a local truncation error of [missing value]. Obtain the predicted value Through the loss function Calculate the prediction error, where Represents the number of time series samples; The adaptive moment estimation optimization algorithm was used for parameter optimization, with the learning rate set to 0.001 and the momentum parameter... =0.9, =0.999, numerical stability constant =10 -8 The iteration count is 1000, and the rate of change of the loss function is less than 10. -8 The optimization is stopped when the loss on the validation set fails to decrease after 10 consecutive iterations to prevent overfitting, thus obtaining the optimal optimization parameters. .

4. The workshop equipment status control system based on the Internet of Things according to claim 3, characterized in that, The process of building a dynamic graph network based on the optimal parameters in S2 to model the spatiotemporal evolution of device states is as follows: Workshop equipment is mapped as graph nodes, and physical connections or data flow associations between equipment are mapped as edges, defining edge weights. Reflecting the degree of influence between nodes, the weight calculation is based on the physical distance between devices and the frequency of data transmission; Based on optimal optimization parameters Dynamically update node attributes The updated formula is: ； in, Indicates the first The set of neighbors of a node. The time step is set to 0.1 seconds; The specific process of spatiotemporal evolution modeling includes: establishing an adjacency matrix based on the device layout, recording node connection relationships to construct a graph structure, and calculating edge weights in real time. ,in The distance between devices. For data transmission frequency, through the spatiotemporal evolution equation Numerical solution is performed, where This is an identity function or linear transformation function used to process the state information of neighboring nodes and set a threshold for vibration frequency changes. =100Hz / s, temperature change threshold =5℃ / s, and when the threshold is exceeded, it is recorded as the optimal mutation state data.

5. A workshop equipment status control system based on the Internet of Things according to claim 1, characterized in that, The process in S3 of introducing a dynamic convolution kernel deformation mechanism based on optimal mutation state data through a deep spatiotemporal convolutional network is as follows: Constructing Deformation Networks The deformation network consists of two convolutional layers and a deformation parameter generator, with the following specific structure: First convolutional layer: kernel size is 3×3, number of channels is 64, stride is 1, padding is 1; The second convolutional layer has a kernel size of 5×5, 128 channels, a stride of 1, and padding of 2. Deformation parameter generator: consists of three fully connected layers with 256, 128, and 6 neurons respectively; The implementation process includes the following steps: processing the input optimal mutation state data through the first convolutional layer. Preliminary feature extraction is performed, among which In terms of time dimension, For spatial dimensions, For the channel dimension, features are further extracted through a second convolutional layer to obtain deep abrupt change state features. Based on deep mutation state characteristics The attention weights of sub-mutation features are calculated using an attention mechanism. The calculation formula is: ； in The learnable parameter matrix; Based on device feature encoding vector Generate deformation parameters : ； Through deformation parameters Adjusting the sampling coordinates of the convolution kernel enables elastic deformation of the convolution kernel, thus optimizing the feature extraction process; The optimal mutation state data is input in tensor form, with its dimensions matching the real-time data stream. Before feature extraction, the optimal mutation state data is concatenated with the real-time device state data of the current time window along the channel dimension, serving as the joint input of the deep spatiotemporal convolutional network. This enables the model to simultaneously perceive instantaneous mutations and continuous states, and finally outputs device control strategy data, including device start / stop commands and operating parameter adjustments, through a fully connected layer.

6. A workshop equipment status control method based on the Internet of Things, using the system described in any one of claims 1-7, characterized in that, The control method comprises the following steps: S1: Deploy multi-source sensors to collect equipment status data in the workshop, and obtain equipment status vectors through normalization preprocessing operations; S2: Construct an intelligent state evolution model based on the device state vector to obtain the optimal optimization parameters; S3: Construct a dynamic graph network based on the optimal parameters to model the spatiotemporal evolution of equipment states and capture the data of the best abrupt change states; S4: Based on the optimal mutation state data, a dynamic convolution kernel deformation mechanism is introduced through a deep spatiotemporal convolutional network to output equipment control strategy data and realize the control of workshop equipment status.

7. The IoT-based intelligent control system for workshop equipment status according to claim 6, characterized in that, The data acquisition module includes multi-source sensors such as vibration sensors, temperature sensors, and current sensors. The data is preprocessed by an edge computing device and then transmitted to the cloud platform.

8. The IoT-based intelligent control system for workshop equipment status according to claim 6, characterized in that, The dynamic convolution kernel deformation mechanism in the control output module is implemented through a deformation network. The deformation network structure consists of two convolutional layers: the first convolutional kernel has a size of 3×3, and the second convolutional kernel has a size of 5×5. The deformation parameter generator adopts a fully connected layer structure.

Images