Remote intelligent real-time online sensing and monitoring system for environment-friendly electricity utilization

Abstract

The invention relates to the technical field of industrial control systems, in particular to a remote intelligent real-time online sensing and monitoring system for environment-friendly power utilization, which comprises a data acquisition unit, a signal processing unit, an output response unit and an automatic updating unit. The data acquisition unit is used for acquiring and standardizing analog signals and digital signals of industrial equipment; the signal processing unit constructs a multi-channel fusion mechanism, generates aggregation state data through dynamic weighting, maps the aggregation state data into equipment association topology, and adjusts the sampling frequency in combination with a dynamic sampling mechanism; the output response unit uses a pre-trained state control model to identify abnormal modes such as shutdown, linkage failure and out-of-limit, and outputs an abnormal level; and the automatic updating unit realizes continuous optimization of the model parameters through a model parameter continuous optimization mechanism. According to the invention, high-precision monitoring and multi-level early warning of a complex industrial state are realized, and the real-time performance and the intelligent level of the system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial control systems, and in particular to a remote intelligent real-time online perception and monitoring system for environmentally friendly electricity use. Background Art

[0002] Given the continuous improvement of environmental protection regulations, the deepening of corporate social responsibility awareness, and the general increase in public environmental awareness, monitoring and managing the energy consumption of industrial pollution control equipment has become a critical issue. Existing solutions primarily rely on system management and manual inspections to identify equipment operating anomalies. This leads to inefficient human resources and technical issues such as inaccurate and timely identification of equipment operating anomalies. Furthermore, the large number of factories and their often remote locations make frequent on-site inspections difficult for managers, further complicating supervision. Furthermore, the current monitoring equipment itself consumes a lot of energy, increasing the overall electricity burden and hindering the goal of environmentally friendly electricity use.

[0003] To meet the above challenges, modern industrial production units urgently need to introduce low-power and efficient automated electricity monitoring systems. These systems can not only accurately monitor the power usage of pollution control equipment and provide abnormal warnings, but also minimize the energy consumption of the monitoring system itself, achieving truly environmentally friendly electricity use and sustainable management, and helping to achieve industrial environmental protection goals.

[0004] Therefore, a remote intelligent real-time online perception and monitoring system for environmentally friendly electricity consumption is proposed. Summary of the Invention

[0005] The purpose of the present invention is to provide a remote intelligent real-time online perception and monitoring system for environmentally friendly electricity consumption, which realizes real-time perception, intelligent analysis and remote control of the electricity consumption behavior of industrial facilities by real-time monitoring of the power consumption and operating status of industrial facilities.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The remote intelligent real-time online sensing and monitoring system for environmentally friendly electricity consumption comprises:

[0008] Data acquisition unit: used to collect multimodal signal data of industrial equipment;

[0009] Signal processing unit: Build a multi-channel fusion mechanism to perform weighted fusion of multimodal signal data and map it into a device-associated topology. Design a dynamic sampling mechanism with a programmable real-time alarm clock and a low-power microcontroller as core energy-saving components. Set a sleep scheduling strategy to achieve periodic wake-up and maintain operating status. Combined with a weak signal detection and processing module, it enables immediate response to key fluctuating signals. Overall sampling control follows the operation control scheduling principle of optimizing energy consumption.

[0010] Output response unit: monitors abnormal modes of industrial equipment and determines abnormality levels through an embedded state control model. Abnormal modes of industrial equipment include shutdown, linkage failure, and limit violation.

[0011] Automatic update unit: Real-time update of state control model parameters through the model parameter continuous optimization mechanism.

[0012] Preferably, the multimodal signal data includes:

[0013] The analog physical quantities of voltage, current, temperature, and vibration are sampled in real time through the sampling frequency set by dynamic sampling. The sampled data is filtered and then formed into an analog signal sampling vector. According to the formula:

[0014] S analog (t)={v(t),i(t),T(t),a(t)}

[0015] Among them, S analog (t) represents the analog signal sampling vector at time t, and v(t), i(t), T(t), and a(t) represent the voltage, current, temperature, and vibration intensity values ​​at time t, respectively. At the same time, binary discrete signals such as the device start / stop status, fault flag, and control logic status are periodically acquired through the digital input interface, taking values ​​of 0 or 1, and constructing a digital signal sampling vector.

[0016] S digital (t)={x1(t),x2(t),...,x n (t)}

[0017] Among them, S digital (t) digital signal sampling vector at time t, x i (t) represents the state of the i-th digital signal at time t, which can be 0 or 1.

[0018] Preferably, the multi-channel fusion mechanism comprises the following specific steps:

[0019] The analog signal is filtered, sampled and quantized in the first channel to obtain a discrete signal. analog (t) Passes through the filter in the first channel to remove the noise in the analog signal. The filtered signal can be expressed as:

[0020]

[0021] Among them, h(t-τ) is the impulse response of the filter, S analog (t) is the original analog signal, S filtered (t) is the filtered signal. Secondly, the continuous time signal is converted into a discrete time signal. According to the signal sampling theorem, the sampling frequency fs Must be higher than the highest frequency of the signal. The discrete signal after sampling is:

[0022] x sampled [n] = x filtered (nT s )

[0023] Among them, x sampled [n] is the discrete signal after sampling, is the sampling period, n is the discrete index of the sampling moment, x filtered is the original discrete signal. Finally, the sampled signal undergoes a quantization process, converting the continuous-amplitude discrete-time signal into discrete values ​​within a finite set of amplitudes. After filtering, sampling, and quantization, the original analog signal is converted into a discrete signal that is convenient for digital processing.

[0024] The digital signal is denoised and normalized in the second channel. First, random noise is removed by wavelet transform. Assume that the filtered signal is x q [n], denoising using a low-pass filter:

[0025]

[0026] Among them, h[k] is the impulse response of the filter, M is the length of the filter, * represents the convolution operation, and y[n] is the denoised signal. Secondly, the signal is normalized by standardization so that the signal is within a certain range. Specifically, the mean and standard deviation of the digital signal are calculated, and then the mean μ is used. y and standard deviation σ y Perform standardization according to the formula:

[0027]

[0028] Among them, z[n] is the normalized signal and y[n] is the denoised signal.

[0029] Assume that the discrete signal output by the first channel is x1, and the output result of the second channel is x2. The dot product attention mechanism is used to calculate the weight, specifically:

[0030]

[0031] Where W is a trainable attention weight parameter, exp(·) is a natural exponential function, and the aggregated state data is calculated based on the weight:

[0032] x fusion =α1·x1+α2·x2

[0033] Among them, α1 and α2 are weights, x1 is the discrete signal output by the first channel, and x2 is the output result of the second channel. fusion Represented as node features in the device association topology, assuming the entire device association topology contains N industrial devices, a graph structure feature G = (V, E) is constructed, where the node set V is constructed by the industrial devices, and E represents the connection relationship between industrial devices. A graph neural network (GCN) is selected as the state control model. By superimposing multiple layers of GCN on the local and global structural features of each node in the device association topology, the model can capture the node's multi-order neighbor information, ultimately outputting a graph embedding feature vector for each node:

[0034] z i =GCN(x fusion ,G)

[0035] Among them, x fusion Comprehensive feature vector, G is the constructed graph structure feature, z i Graph embedding feature vector.

[0036] Preferably, the dynamic sampling mechanism comprises the following specific steps:

[0037] The low-power microcontroller starts the initialization program and loads the system configuration parameters, including sampling channels, wake-up cycles, and signal thresholds;

[0038] Configure the programmable real-time alarm clock RTC as the system's main wake-up timer and set the initial sampling period;

[0039] Use ultra-low power ADC module to enable weak signal sensing;

[0040] The controller enters sleep mode and turns off the power of all peripherals except the RTC and interrupt channels to minimize system power consumption;

[0041] When the RTC wake-up time is reached, the controller wakes up, starts the data acquisition unit, samples from each modal sensor, and pre-processes the sampled data to determine whether there is fluctuation;

[0042] If the data is stable, a longer RTC wake-up cycle is reset and the system goes back to sleep. If a slight fluctuation is detected but does not exceed the set threshold, the next wake-up cycle is shortened and the trend is monitored intensively.

[0043] During the main control's sleep period, the weak signal sensing module keeps running. Once a signal mutation exceeds the threshold, an interrupt is triggered to wake up the main control. After waking up, it immediately enters the high-frequency sampling mode and locks the signal characteristics of the current time period.

[0044] The system records the current sampling frequency and device load status in real time, evaluates the energy efficiency ratio, and updates the policy parameters to achieve long-term optimal energy consumption. If there are no abnormal fluctuations for multiple consecutive cycles, the system automatically returns to the lowest power consumption sampling frequency.

[0045] Preferably, the specific structure of the state control model is:

[0046] A graph neural network is selected as the state control model. The graph structure is constructed based on the energy consumption correlation and collaborative operation relationship between industrial equipment nodes. The connection weights between devices are represented in the form of a weighted adjacency matrix. The whole model consists of a logic analysis layer, a state synthesis layer, and a state determination layer. It is used to extract local and global power consumption patterns and perform state identification. Specifically, it includes:

[0047] The logic analysis layer is used to extract the structural and attribute information between each device node and its neighboring nodes in the industrial field. Through multi-layer graph convolution operations, node features are propagated and updated layer by layer. The core operations of each convolution layer are as follows:

[0048]

[0049] in, is the adjacency matrix after adding the self-loop, which represents the connection relationship between devices; for The degree matrix of (l) is the node feature matrix of the lth layer; W (l) is the training learnable weight matrix; σ represents the activation function, which realizes the embedded expression of the correlation between the operating states of industrial equipment.

[0050] The state synthesis layer performs neighborhood aggregation on the features of each device node, introduces local context information, and realizes the collaborative expression of power usage patterns between devices. The aggregation results enhance the node's discrimination ability.

[0051] The state determination layer outputs the state recognition of the final fused node features. Through the fully connected layer and the softmax layer, the current embedded features of the device are mapped to the multi-classification output label y according to the formula:

[0052] y=softmax(W c h v +b c )

[0053] Where W c 、b c is a trainable parameter, h v The final node representation.

[0054] Preferably, the current status of the industrial equipment includes:

[0055] The current status of the industrial equipment includes shutdown, linkage failure and over-limit.

[0056] Shutdown refers to the abnormal cessation of industrial equipment operation, which is usually manifested by the analog signal suddenly dropping to zero or maintaining an extremely low value in a short period of time. The analog signal is monitored over multiple consecutive time steps to see if it meets the following conditions:

[0057] P(t)=ε p , and duration Δt>T min

[0058] Where P(t) represents the power value at a certain moment, ε p is the power threshold, Δt is the abnormal duration, and if the conditions are met, it is determined to be a shutdown abnormality.

[0059] In industrial processes, multiple devices often form a linkage control chain. The startup of an upstream device should automatically trigger the operation of downstream devices. If the system detects that there is an expected device state that is not triggered in the linkage chain, it is determined to be a linkage failure. This can be defined by analyzing the timing dependency graph between devices and the mutual consistency of their states:

[0060] S i (t)=Active,S i (t+δ)≠Active

[0061] This indicates that the operation of device i fails to effectively link device j, thus triggering a linkage failure warning.

[0062] When the voltage, current, temperature and other key indicators of the equipment during operation exceed the preset physical or safety threshold, the system will determine it as an over-limit abnormality. The system further evaluates the magnitude and duration of the over-limit abnormality as the basis for judging the abnormality level:

[0063] Normal: no abnormalities;

[0064] Slight abnormality: The indicator just exceeds the normal fluctuation range, but the duration is short and the amplitude is small, and there is no obvious impact on system operation;

[0065] Moderate abnormality: The indicator deviates significantly or the linkage abnormality occurs repeatedly, which may have a certain impact on the equipment operation efficiency or product quality, but will not affect the system safety;

[0066] Serious anomalies: These include continuous shutdowns, significant electrical indicator violations, and complete failure of key link linkages, which seriously threaten the safe operation of the system and require immediate alarms and manual intervention.

[0067] Preferably, the specific process of the model parameter continuous optimization mechanism is:

[0068] To ensure the system's continuous adaptability to complex operating conditions and the ability to update models, a continuous optimization mechanism for model parameters was designed in conjunction with the state control model. During the initial system deployment, a large amount of historical operating status and power monitoring data from the factory was collected. This data was then used to generate a training dataset through graphical modeling and manual annotation. This dataset was then used to train the state control model offline, generating an initial model.

[0069] During system operation, sensors and power devices continuously collect multimodal signal data. The system dynamically accesses the data stream using a sliding time window mechanism. To determine signal stability, the system calculates the variance within the time window and removes signals with feature loss ratios exceeding a threshold η. Qualified signals are converted into a graph structure and, along with corresponding state labels, form new data for the continuous optimization mechanism of model parameters. Using a small-batch fine-tuning strategy, these data are grouped and fed into the state control model for local updates. During this update, only the parameters of the logic analysis layer and the state synthesis layer are adjusted.

[0070] To avoid overfitting, an early stopping mechanism is introduced during training. If the model's accuracy on the validation set does not significantly improve over multiple training cycles, the model is considered to have converged and the update process ends. The updated graph neural network model is then immediately used for real-time prediction tasks, identifying the power usage behavior and state of industrial equipment. Combined with the system's remote feedback mechanism, it triggers early warnings and forms a closed-loop "data-model-control" process, enabling intelligent, highly robust, real-time online monitoring and management of industrial power consumption.

[0071] Compared with the prior art, the present invention has the following beneficial effects:

[0072] 1. The present invention proposes a dynamic adjustment mechanism for adjusting the sampling frequency when collecting multimodal signal data. This mechanism adjusts the sampling frequency in real time based on the operating status of the industrial equipment and the degree of environmental disturbance to balance the accuracy of data collection and the optimization of system energy consumption. By adopting a programmable real-time alarm clock and a low-power microcontroller, the device can be periodically awakened and switched to sleep. In combination with weak signal sensing technology, it can respond immediately to key fluctuation signals, thereby effectively reducing the overall energy consumption of the monitoring system and improving environmentally friendly electricity efficiency. This mechanism not only ensures high-precision monitoring of abnormal conditions of industrial equipment, but also significantly extends the battery life of the equipment. It is suitable for the remote intelligent monitoring needs of large-scale distributed industrial sites.

[0073] 2. The present invention uses a state control model to analyze multimodal signal data in real time, which has higher accuracy and flexibility than traditional anomaly detection technologies based on rules or statistical methods. Traditional methods usually rely on fixed thresholds or predefined rules, which may not be able to effectively handle nonlinear relationships and dynamic changes between complex devices. The state control model uses a graph structure to model the dependencies between devices, which can automatically learn and capture complex patterns and anomalies. Compared with traditional machine learning methods, the state control model can not only discover potential correlations in multidimensional data, but also automatically optimize its detection capabilities through end-to-end learning, making anomaly detection more adaptive and real-time.

[0074] 3. This invention proposes a continuous model parameter optimization mechanism, enabling the model to continuously absorb new monitoring data and perform self-optimization. This mechanism automatically updates existing model parameters through a small-batch fine-tuning strategy, rapidly adapting to emerging changes in industrial equipment status while preserving historical knowledge. Compared to traditional full-scale retraining methods, this mechanism significantly reduces computing resource consumption and training time, improving the model's adaptability and real-time performance in dynamic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 This is a flow chart of a remote intelligent real-time online perception and monitoring system for environmentally friendly electricity use proposed in an embodiment of the present invention;

[0076] Figure 2 This is a structural diagram of a remote intelligent real-time online perception and monitoring system for environmentally friendly electricity use proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0077] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0078] During the operation of industrial pollution control facilities, their energy consumption and operating status are directly related to pollutant treatment efficiency, equipment lifespan, and overall environmental performance. Abnormal operation of the facilities can not only lead to excessive pollutant emissions but also cause equipment failure, resulting in production interruptions or environmental pollution incidents. To ensure the stable and efficient operation of industrial pollution control facilities, real-time collection and monitoring of key operating parameters are particularly important. By continuously acquiring and analyzing this operating data, potential operational anomalies can be promptly identified, operational control strategies can be optimized, and reliable data support can be provided for subsequent anomaly warnings, fault diagnosis, and operation and maintenance decisions. This effectively reduces equipment failure rates, improves environmental compliance rates, and enhances the overall safety of system operations.

[0079] This invention proposes a remote, intelligent, real-time, online sensing and monitoring system for environmentally friendly electricity consumption. This system collects and analyzes the energy consumption characteristics of facilities during operation in real time, improving the accuracy and response speed of anomaly identification. To illustrate the effectiveness of the method and system, this invention will be described in detail with reference to the accompanying drawings and the following two examples.

[0080] Example 1

[0081] Traditional power monitoring and environmental governance approaches face multiple challenges in the operation of industrial pollution control facilities at large chemical companies, particularly in monitoring high-power, high-frequency equipment. Existing solutions primarily rely on system management and manual inspections to identify equipment anomalies. This leads to inefficient human resources and inaccurate and timely identification of equipment anomalies. Furthermore, the large number of factories and their often remote locations make frequent on-site inspections difficult for managers, further complicating oversight. Furthermore, the high energy consumption of current monitoring equipment increases the overall power burden and hinders the achievement of environmentally friendly electricity goals.

[0082] In order to overcome the above limitations of the existing industrial equipment monitoring system, taking the grinding fan of chemical enterprise A as an example, this embodiment elaborates on a remote intelligent real-time online sensing monitoring system for environmentally friendly electricity consumption, such as Figure 1 Shown, including:

[0083] Professional-grade sensors are placed on the monitoring equipment to collect multimodal signal data during the operation of the equipment in real time and perform standardized processing (S1). The multimodal signals are weighted and fused through a multi-channel fusion mechanism to generate aggregated state data (S2). The aggregated state data is mapped to the device association topology, and the signal sampling frequency is adjusted through a dynamic sampling mechanism (S3). Then, a pre-trained state control model is used to identify abnormal states of the equipment and output the abnormality level (S4). Finally, the parameters of the state control model are updated in real time through a continuous model parameter optimization mechanism (S5).

[0084] Specifically, the multimodal signal data acquisition process of the wind turbine includes:

[0085] Acquiring analog signals involves deploying high-precision sensors at key locations, such as the fan bearings, motor, cooling unit, and air duct outlet, to collect continuous physical quantities such as current, voltage, temperature, vibration, and noise. For example, the analog signal sampling frequency is set to 2000Hz, with 16-bit accuracy. This is used to reflect load fluctuations, abnormal temperature rises, and shaft vibration characteristics during fan operation. Initial filtering and amplification are then performed using the industrial signal conditioning module.

[0086] Acquiring digital signals involves collecting discrete switching values ​​from the fan control logic through the industrial control system interface, including start / stop signals, interlocking status, air inlet and outlet valve openings, and overload alarm triggering flags. For example, the acquisition cycle is 1 second to ensure a complete record of the switching logic timing during fan operation.

[0087] Constructing a unified signal standardization dataset involves uniformly encoding, normalizing, and time-aligning the aforementioned analog and digital signals through an acquisition gateway, generating multimodal signal data with consistent structure and continuous temporal sequence. This data structure supports efficient integration with the input format of graph neural network models.

[0088] In this way, by uniformly aligning the timing of the collected analog and digital signals, normalizing their amplitudes, removing outliers, and filtering noise, the consistency of the input signals in terms of time base, the uniformity of the data format, and the integrity of the signal characteristics are ensured. Analog and digital signals are structured and mapped in a unified data channel, providing high-quality multimodal input data for subsequent multi-channel fusion, graph structure modeling, and model inference. This normalization process significantly improves the system's data robustness and monitoring stability in complex industrial field environments, effectively avoiding misjudgments caused by missing sensor data, bit errors, or asynchronous sampling.

[0089] Furthermore, the multi-channel fusion mechanism has the following specific steps:

[0090] The multi-channel fusion mechanism includes the first channel, the second channel and the third channel;

[0091] The first channel is used to provide the characteristics of continuous control variables when filtering, sampling and quantizing the analog signal. Specifically, the collected analog signal is first processed by a three-stage Butterworth low-pass filter with a filter cutoff frequency set to 50Hz to effectively filter out high-frequency noise and electromagnetic interference, thereby improving signal quality.

[0092] The filtered analog signal is sampled at a 100Hz sampling frequency with a sampling period of 0.01 seconds, ensuring high-resolution capture of the wind turbine's dynamic state. The sampled signal is then quantized using a 12-bit analog-to-digital converter with a resolution of approximately 2.44mV, converting the continuous analog signal into a digital signal and providing the foundational data for subsequent digital processing. Ultimately, the first channel outputs a 256-dimensional continuous control variable feature vector.

[0093] The second channel processes digital signals, collecting six digital signals from the wind turbine, including equipment start / stop status, linkage operation flags, and fault alarm signals. First, multi-layer denoising is performed using wavelet transform to effectively suppress random interference and communication errors. The denoised signals are then normalized to a value between 0 and 1 based on statistical parameters with a mean of 0.48 and a standard deviation of 0.15, ensuring numerical consistency across different signals. The second channel ultimately outputs a 128-dimensional discrete control state feature vector.

[0094] The third channel fuses the feature vectors output by the first and second channels, automatically calculates the weight distribution of the two features using the dot product attention mechanism, and generates a 512-dimensional comprehensive feature vector to enhance the complementarity of multimodal signal data;

[0095] This aggregated state data is then mapped to node features in the device topology. Five key industrial equipment nodes are selected in the system to construct a device association graph with eight connecting edges, reflecting the control and physical coupling relationships between the devices. A three-layer state control model is used to comprehensively extract local and global structural information from each node in the graph, reducing the feature dimensions from 512 to 128 and 64 layer by layer, ultimately outputting a 32-dimensional node embedding feature.

[0096] Through the attention-weighted fusion mechanism of the third channel, the system can dynamically coordinate the continuous physical quantity characteristics of the first channel and the discrete control state characteristics of the second channel, integrate the complementary advantages of the two into a unified comprehensive representation, and map the representation into a device state association graph structure, thereby capturing the local details and global operating modes of the signal in the same model, achieving deep perception and accurate prediction of the operating status of industrial equipment.

[0097] Furthermore, the specific process of the dynamic sampling mechanism is as follows:

[0098] Based on the operating status of industrial equipment and the degree of environmental disturbance, the sampling frequency is adjusted in real time, with the specific sampling frequency range set to 10Hz to 1kHz;

[0099] When the device is in a stable operating state, the sampling frequency is automatically reduced to 10Hz to maximize energy saving;

[0100] When the fluctuation index is within a stable range below 0.1 and remains in this state for 5 consecutive minutes, the sampling frequency is gradually reduced from 50Hz to 10Hz to reduce the data volume and processing burden. When operating fluctuations or abnormal signals are monitored, the sampling frequency is quickly increased to the range of 500Hz to 1kHz to ensure the collection of key details.

[0101] Using a low-power microcontroller (less than 5mW) and a programmable real-time alarm clock, it implements intelligent sleep scheduling of the device by periodically waking up (with a period range of 1 to 10 seconds) and switching between sleep and wake-up.

[0102] Combined with a weak signal sensing module with a sensitivity of -90dBm, it enables immediate response to critical fluctuation signals. The overall mechanism adheres to the principle of optimal energy consumption, reducing the average power consumption of the monitoring system by over 30%, effectively extending device battery life and improving environmentally friendly power efficiency.

[0103] By introducing a dynamic sampling mechanism centered around a programmable real-time alarm clock and a low-power microcontroller, combined with weak signal sensing capabilities, this system achieves immediate response to critical fluctuations in industrial equipment, effectively enhancing the system's adaptive sensing capabilities under both stable and fluctuating operating conditions. Based on an intelligent scheduling strategy for optimal energy consumption, this mechanism dynamically adjusts the sampling frequency, significantly reducing system energy consumption while ensuring monitoring accuracy. This reduces the burden of ineffective data collection and processing, enhancing the overall system's intelligence, energy efficiency, and robustness, and meeting the demands of environmentally friendly electricity applications.

[0104] Furthermore, the specific structure of the state control model is:

[0105] The model was pre-trained offline using 1,200 annotated wind turbine fault and normal operation data collected over the past six months (each data record contains time series features of 32 device nodes).

[0106] The model consists of three logical analysis layers, two state synthesis layers, and one state determination layer. The logical analysis layers have 64, 32, and 16 channels, respectively. The state synthesis layer uses a mix of average and max pooling, and the state determination layer is a 16-by-4 fully connected network (with four output categories: normal, slightly abnormal, moderately abnormal, and severely abnormal).

[0107] In the logic analysis layer, the input 512-dimensional comprehensive feature vector is first fed into the first layer of graph convolution, which uses a 3×3 convolution kernel to extract the local structural information of the node and its neighbors, and outputs a 64-dimensional node representation. Subsequently, the second layer of graph convolution takes the 64-dimensional representation as input, further extracts deeper topological features, outputs 32-dimensional features, and enhances the nonlinear expression through ReLU activation. The third layer of graph convolution receives the 32-dimensional features and outputs a 16-dimensional representation. LayerNorm normalization is applied after each convolution to ensure the consistency of feature distribution of different nodes.

[0108] In the state aggregation layer, the system uses both neighbor feature averaging and neighbor feature maximization in parallel, concatenating the resulting vectors into a 32-dimensional context vector. After each aggregation, some features are randomly discarded at a ratio of 0.1 to prevent overfitting. Finally, the classification layer feeds this 32-dimensional context vector into a 16-by-4 fully connected network. A softmax is used to output four-class classification probabilities. Training is performed for 50 epochs using the cross-entropy loss function and the Adam optimizer (initial learning rate 0.001, batch size 64) to accurately classify the operating status of industrial equipment.

[0109] By mapping the comprehensive feature vector into a graph structure after node feature fusion, the system innovatively captures the relationship between the local node and its multi-order neighbors in the network through multi-layer graph convolution layers, and fuses the average and maximum features through a parallel aggregation strategy to achieve accurate modeling of control dependencies between devices; this structured representation is then directly used for multi-level fault classification in the state judgment layer, allowing the model to take into account both local abnormality details and global coupling patterns within a single architecture, significantly improving the ability to recognize complex abnormalities such as shutdowns, linkage failures, and over-limits.

[0110] Furthermore, the current status of the industrial equipment is specifically:

[0111] When the predicted shutdown probability of the grinding fan exceeds 0.9 and the measured current is continuously lower than 5A for more than 3 seconds, the system determines it as a "shutdown" abnormality;

[0112] If the startup commands of two adjacent devices fail to be triggered within 2 seconds, it is identified as "linkage failure";

[0113] When the voltage or vibration amplitude exceeds the normal value by ±10% and lasts for 5 seconds, it is judged as "out of limit". Based on the above three types of abnormalities, the system generates the corresponding abnormal level according to the preset four-level response strategy:

[0114] Normal: All indicators are within the threshold range;

[0115] Minor abnormality: If the deviation from the threshold is ≤10% and lasts for <5 seconds, only a log alarm will be issued;

[0116] Moderate abnormality: If the deviation from the threshold is 10-20% or lasts for 5-15 seconds, local load reduction or standby fan linkage will be initiated;

[0117] Severe abnormality: If the deviation from the threshold is greater than 20% or lasts for more than 15 seconds, the entire plant will be shut down for protection immediately, and an emergency alarm will be pushed through the remote alarm center.

[0118] The graph neural network model is used to evaluate the equipment status in real time and compare it with the preset threshold. Once it is determined to be a shutdown, linkage failure or over-limit abnormality, the corresponding control action (such as shutdown protection, linkage retry or load reduction isolation) is immediately triggered and a remote alarm is issued. This hierarchical response not only ensures the rapid handling of critical faults, but also balances system stability and production continuity by fine-tuning the control strategy.

[0119] Furthermore, the specific steps of the model parameter continuous optimization mechanism are as follows:

[0120] During the system deployment phase, a multimodal signal dataset containing labels for normal operating conditions and typical faults (such as shutdown, linkage failure, and over-limit) was screened from historical operation records and maintenance logs. The model was trained offline to generate an initial state control model and verify its baseline performance on the wind turbine state classification task.

[0121] After the system goes online, the data acquisition unit continuously receives multimodal signals from the wind turbine. The online data preprocessing module aggregates new sampled data every minute and selects qualified new samples based on the conditions that the vibration amplitude variance within 60 consecutive seconds is less than the preset threshold and that no key switching values ​​are lost.

[0122] New samples are fed into the incremental training process in batches of 32, and only the parameters of the first two layers of the state control model, the logic analysis layer and the state synthesis layer, are fine-tuned to quickly adapt to the state changes of the wind turbine under different load, temperature or environmental conditions.

[0123] After each small-batch fine-tuning, the system evaluates the model accuracy on a specially reserved validation set (the most recent 200 samples). If the accuracy improvement is less than 0.5% over five consecutive fine-tuning cycles, the model is automatically considered to have converged and no further updates are performed in this round.

[0124] The updated model was immediately deployed back to the monitoring platform for subsequent monitoring of X-grinding fan power consumption and intelligent early warning.

[0125] By continuously receiving high-quality operating samples from the wind turbine's multimodal signals and screening them for timing stability and integrity, they are dynamically incorporated into the pre-trained state control model. Through small-batch fine-tuning, the local parameters of the logic analysis layer and the state synthesis layer are quickly updated to ensure that the model can adapt to the state changes of the wind turbine under different loads and environments in real time. After monitoring the performance improvement on the validation set and automatically determining convergence, it seamlessly switches to the latest model for subsequent power consumption behavior monitoring and intelligent control, so as to continuously maintain high-precision identification and response capabilities to abnormal states of the wind turbine.

[0126] Example 2

[0127] In Example 1, the method of the present invention realizes real-time monitoring of the grinding fan of Enterprise A. In the embodiment of this application, the remote intelligent real-time online sensing monitoring system for environmentally friendly electricity proposed by the present invention is applied to the dust removal equipment of the brushing machine of Textile Enterprise B to monitor in real time whether there is any abnormality in the pollution control facility. Figure 2 ,The monitoring system includes a data acquisition unit, a control ,processing unit, an output response unit and an automatic ,update unit.

[0128] Furthermore, high-precision sensors are arranged at key locations of the brushing machine and dust removal equipment to collect real-time analog signal data of current, voltage, temperature and vibration, as well as digital signal data of equipment start and stop status, linkage operation signs and fault alarm signals.

[0129] The collected analog signals are then filtered, sampled, and quantized, while the digital signals are denoised and normalized. Subsequently, the two feature vectors are fused using a dot-product attention mechanism to generate aggregated state data, enhancing the complementarity of the multimodal signal data.

[0130] Furthermore, the comprehensive feature vectors are mapped to node features in a graph structure, constructing a device association topology that reflects the control and physical coupling relationships between devices. Through three logical analysis layers, local and global structural information is extracted for each node in the graph, enabling in-depth analysis of device operating status.

[0131] Furthermore, the system calculates the fluctuation index every 10 seconds and automatically adjusts the sampling frequency according to the equipment operating status, ensuring high-resolution information at critical moments while saving storage and computing resources during steady-state operation.

[0132] Furthermore, the state control model is used to evaluate the equipment status in real time and compare it with the preset threshold. Once it is determined to be a shutdown, linkage failure or over-limit abnormality, the corresponding control action is immediately triggered and a remote alarm is issued.

[0133] To further verify the improvements of the proposed remote intelligent real-time online sensing and monitoring system for environmentally friendly electricity use compared to traditional monitoring systems in terms of fault identification, response time, equipment operation stability, and energy consumption efficiency, the following experiments were conducted:

[0134] Table 1 Efficiency comparison with traditional monitoring system

[0135] Traditional monitoring system This monitoring system Improvement ratio Fault identification rate (%) 75 92 17％ Average response time (seconds) 30 12 -60％ Equipment downtime (times) 3 1 -66.67％ Energy consumption efficiency (kWh) 1500 1300 -13.33％ Anomaly detection rate (%) 78 95 17％

[0136] The experimental results shown in Table 1 show that the remote intelligent real-time online perception monitoring system for environmentally friendly electricity consumption of the embodiment of the present application has significant advantages over traditional monitoring systems. Through the real-time acquisition and intelligent processing of multimodal signals, the system reduces noise interference and improves the accuracy of fault warnings, especially in the identification of abnormal power outages, linkage failures and over-limit situations. In addition, the system enhances real-time performance and control accuracy by dynamically adjusting the sampling frequency, ensuring stable operation of the equipment and reducing energy consumption. In general, this system optimizes the data acquisition, processing and control decision-making links, improves the stability and energy efficiency of equipment operation, and verifies its innovative advantages in improving monitoring effects and response speed.

[0137] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A remote intelligent real-time online sensing and monitoring system for environmentally friendly electricity consumption, characterized in that: include: Data acquisition unit: used to collect multimodal signal data of industrial equipment; Signal processing unit: Build a multi-channel fusion mechanism to perform weighted fusion of multimodal signal data and map it into a device-associated topology. Design a dynamic sampling mechanism with a programmable real-time alarm clock and a low-power microcontroller as core energy-saving components. Set a sleep scheduling strategy to achieve periodic wake-up and maintain operating status. Combined with a weak signal detection and processing module, it enables immediate response to key fluctuating signals. Overall sampling control follows the operation control scheduling principle of optimizing energy consumption. Output response unit: monitors abnormal modes of industrial equipment and determines abnormality levels through an embedded state control model. Abnormal modes of industrial equipment include shutdown, linkage failure, and limit violation. Automatic update unit: Real-time update of state control model parameters through the model parameter continuous optimization mechanism.

2. The remote intelligent real-time online sensing and monitoring system for environmentally friendly electricity use according to claim 1 is characterized in that: The multimodal signal data is specifically: Analog and digital signals are acquired in a programmed manner by a distributed industrial sensor array. The analog signal is used to reflect the key control parameters during the operation of industrial equipment, and the digital signal is a discrete switching quantity that indicates the start and stop status of the equipment.

3. The remote intelligent real-time online sensing and monitoring system for environmentally friendly electricity use according to claim 1 is characterized in that: The multi-channel fusion mechanism is specifically as follows: The multi-channel fusion mechanism includes the first channel, the second channel and the third channel; The first channel is used to provide state parameters of continuous process control variables when filtering, sampling and quantizing analog signals; The second channel is used to provide operation indication information of discrete logic states when performing denoising and standardization processing on digital signals; The third channel is used to perform dynamic weighted fusion processing on the data processed by the first channel and the second channel, output aggregated state data, and map the aggregated state data into a device association topology. Local and global control logic information is extracted through the state control model to generate a multi-dimensional state description of the operation of industrial equipment.

4. The remote intelligent real-time online sensing and monitoring system for environmentally friendly electricity use according to claim 1 is characterized in that: The dynamic sampling mechanism is specifically as follows: Based on the degree of environmental disturbance and fluctuations in the operating status of industrial equipment, the sampling frequency of multimodal signal data is dynamically adjusted according to preset thresholds; When the state of the industrial equipment is in a fluctuation range, the system wakes up the data acquisition unit and increases the sampling frequency through a programmable real-time alarm clock and a low-power microcontroller; When the device status is in a stable range, the sampling frequency is reduced and the system enters a dormant state to achieve energy-saving operation; Integrate the optimization decision principle of optimal energy consumption and adjust the sampling frequency in real time.

5. The remote intelligent real-time online sensing and monitoring system for environmentally friendly electricity use according to claim 3 is characterized in that: The state control model is specifically: Initial construction is performed through historical operation data to obtain a pre-trained equipment operation logic model. The equipment operation logic model includes a logic analysis layer, a state synthesis layer, and a state determination layer, where: The logic analysis layer performs multi-layer correlation analysis on the device correlation topology data to extract the correlation information of the internal components of the industrial equipment and the control dependency relationship between the devices; The state integration layer integrates the state information of adjacent nodes to achieve correlation evaluation of cross-node operation status; The status determination layer diagnoses and rates abnormal conditions of industrial equipment based on aggregated status data.

6. The remote intelligent real-time online sensing and monitoring system for environmentally friendly electricity use according to claim 5 is characterized in that: The current status of the industrial equipment includes: The current status of the industrial equipment includes normal operation, shutdown, linkage failure and over-limit, and generates corresponding abnormality levels based on preset thresholds and abnormality levels. The abnormality levels include normal, slight abnormality, moderate abnormality or severe abnormality, and directly trigger corresponding automatic control actions in combination with thresholds to issue remote alarm instructions.

7. The remote intelligent real-time online sensing and monitoring system for environmentally friendly electricity use according to claim 1 is characterized in that: The specific process of the model parameter continuous optimization mechanism is as follows: Continuously acquire real-time multimodal signal data, dynamically screen out new samples that meet the requirements of time series stability and state parameter integrity, and achieve continuous optimization of model parameters; The model parameter continuous optimization mechanism receives new samples and adjusts the parameters of the local control logic in the equipment operation logic model based on a small batch parameter correction strategy based on recent data.

Images