Intelligent monitoring system oriented to complex working conditions and reinforcement learning optimization method thereof

Abstract

The invention relates to the field of industrial automation control, in particular to an intelligent monitoring system oriented to complex working conditions and a reinforcement learning optimization method thereof. Comprising a data acquisition module used for acquiring operation data of industrial equipment under complex working conditions; the data preprocessing module is used for preprocessing the operation data acquired by the data acquisition module and then acquiring standard data; the feature extraction module is used for acquiring key feature vectors capable of reflecting the operation state of the industrial equipment; the intelligent monitoring model building module is used for building an intelligent monitoring model, and the intelligent monitoring model takes the key feature vector as input and outputs a prediction result of the operation state of the industrial equipment after interactive learning; and the alarm module is used for judging whether the industrial equipment is in an abnormal state according to the prediction result, and sending an alarm signal if the industrial equipment is in the abnormal state. According to the method, the problems of low disk monitoring efficiency, poor accuracy, insufficient model adaptability, incomplete alarm mechanism and the like under complex working conditions in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of industrial automation control, and in particular to an intelligent monitoring system for complex working conditions and its reinforcement learning optimization method. Background Technology

[0002] In modern industrial production, with the continuous expansion of production scale and the increasing complexity of production processes, extremely high demands are placed on the operation monitoring of production equipment. Traditional monitoring methods mainly rely on manual labor, requiring operators to deal with a large number of operation monitoring screens and massive monitoring parameter points. For example, in the field of thermal power generation, unit start-ups and shutdowns are frequent, load rates and operating hours fluctuate greatly, and operating conditions are complex and changeable. Operators must simultaneously analyze hundreds of operation monitoring screens and tens of thousands of monitoring parameter points. Analyzing such a large number of parameters relying solely on manual labor and experience is an enormous workload and highly prone to oversights.

[0003] While some existing intelligent monitoring systems incorporate some automation technologies, they suffer from numerous shortcomings. Firstly, their data processing and analysis capabilities are limited under complex operating conditions, making it difficult to accurately extract key information. For example, for systems with complex coupling relationships between parameters, existing data feature extraction methods cannot effectively uncover these potential relationships, leading to inaccurate assessments of equipment operating status. Secondly, existing intelligent monitoring systems have deficiencies in model building and optimization. Most systems use overly simplistic models that cannot adapt well to the dynamic changes of complex operating conditions and lack effective optimization methods, resulting in predictive accuracy and reliability that fail to meet practical needs. Furthermore, existing systems are inadequate in their alarm mechanisms, often relying solely on simple threshold judgments, which easily lead to false alarms or missed alarms, failing to provide timely and accurate early warning information to operators. Therefore, there is an urgent need to develop an intelligent monitoring system and its optimization methods that can effectively cope with complex operating conditions and improve the automation, intelligence, comprehensiveness, and accuracy of monitoring. Summary of the Invention

[0004] In order to overcome the above-mentioned technical defects, the purpose of this invention is to provide an intelligent monitoring system for complex working conditions and its reinforcement learning optimization method, so as to solve the problems of low monitoring efficiency, poor accuracy, insufficient model adaptability and imperfect alarm mechanism in the prior art under complex working conditions.

[0005] This invention discloses an intelligent monitoring system for complex working conditions, characterized in that it includes:

[0006] The data acquisition module is used to collect operating data of industrial equipment under complex working conditions;

[0007] The data preprocessing module, connected to the data acquisition module, is used to preprocess the running data acquired by the data acquisition module to obtain standard data.

[0008] The feature extraction module is connected to the data preprocessing module and uses a deep belief network algorithm to extract features from standard data in order to obtain key feature vectors that can reflect the operating status of industrial equipment.

[0009] The intelligent monitoring model construction module uses reinforcement learning algorithms to build an intelligent monitoring model. The intelligent monitoring model takes key feature vectors as input and outputs the prediction results of the operating status of industrial equipment after interactive learning.

[0010] The alarm module, connected to the intelligent monitoring model construction module, is used to determine whether the industrial equipment is in an abnormal state based on the prediction results. If the industrial equipment is in an abnormal state, an alarm signal is issued.

[0011] Preferably, it also includes a high-precision clock module connected to the data acquisition module. The high-precision clock module provides a time synchronization reference for the data acquisition module. When the data acquisition module collects operating data through different types of sensors, the high-precision clock module timestamps the operating data collected by different types of sensors, and the synchronization error is <1ms.

[0012] The data acquisition module adaptively adjusts the sampling rate based on the dynamic changes in the operating conditions of the industrial equipment. The formula for calculating the adaptive sampling rate is as follows:

[0013]

[0014] in, The current sampling rate, The sampling rate at the previous time step. This is an adjustment factor (with a value range of 0.1 to 0.5). The difference between the current operating condition and the operating condition at the previous moment (obtained by calculating the weighted sum of the change rates of key parameters, with a value range of 0 to 1).

[0015] Preferably, the preprocessing in the data preprocessing module includes cleaning, denoising, and normalizing the running data to obtain standard data; wherein,

[0016] When cleaning the running data, the data preprocessing module uses a median filtering algorithm based on a sliding window to remove noisy data. The cleaning calculation formula is as follows:

[0017]

[0018] in, The filtered value. is the initial value, and n is the half width of the sliding window.

[0019] Preferably, during feature extraction, the deep belief network algorithm adjusts the network parameters by minimizing the inter-layer reconstruction error. The formula for calculating the inter-layer reconstruction error is:

[0020]

[0021] in, Let M be the reconstruction error of the l-th layer, and M be the number of neurons in the l-th layer. This represents the actual output value of the j-th neuron in the l-th layer. This is the reconstructed output value of the j-th neuron in layer l.

[0022] Preferably, the reinforcement learning algorithm employs a deep Q-network algorithm, and the Q-value update formula for the deep Q-network algorithm is:

[0023]

[0024] in, In the state Next collection action Q value, For learning rate, In the state Next collection action The reward received later As a discount factor, For the next state, For the next action.

[0025] Preferably, when determining whether industrial equipment is in an abnormal state, the alarm module makes the judgment based on a combination of dynamic adjustment threshold for prediction deviation and trend analysis. The formula for calculating the dynamic adjustment threshold is as follows:

[0026]

[0027] in, The dynamic threshold of parameter p at the current time. Let p be the initial baseline threshold. This is the deviation adjustment factor (range: 0.3~0.8). For parameter p in the past Prediction bias at time ( , Let p be the actual value of parameter p at time q. (where p is the predicted value of parameter, and w is the sliding window length); when the value of parameter p in the prediction result exceeds the dynamic threshold... If the trend of parameter value change meets the preset abnormal conditions within z consecutive sampling periods, the industrial equipment is determined to be in an abnormal state.

[0028] The second objective of this invention is to provide a reinforcement learning optimization method for an intelligent monitoring system under complex working conditions, comprising the following steps:

[0029] Step S100: Collect historical operating data of industrial equipment under complex working conditions and label it, labeling the industrial equipment with tags under different operating states;

[0030] Step S200: Divide the labeled running data into a training set and a test set;

[0031] Step S300: Train the intelligent monitoring model based on the training set, and update the parameters of the intelligent monitoring model using the deep Q-network algorithm during the training process;

[0032] Step S400: Test the intelligent monitoring model with updated parameters based on the test set to evaluate the performance indicators of the intelligent monitoring model with updated parameters and obtain the test results;

[0033] Step S500: Based on the test results, adjust the hyperparameters of the intelligent monitoring model, and train and test again until the performance of the intelligent monitoring model reaches the set standard.

[0034] Preferably, during annotation in step S100, the semi-supervised active annotation strategy based on uncertainty quantification includes the following steps:

[0035] Step S110: Train an initial classification model based on the labeled data, and calculate the prediction entropy value of the unlabeled data. The calculation formula is:

[0036]

[0037] in, Let be the predicted entropy value of data d, and K be the total number of categories. Predict the probability that data d belongs to category k for the model;

[0038] Step S120: Obtain labeled data with predicted entropy values ​​greater than a preset threshold, and have them labeled by domain experts;

[0039] Step S130: Add the newly labeled data to the training set and iteratively update the initial classification model until the average prediction entropy of the unlabeled data is lower than the convergence threshold.

[0040] Preferably, in step S400, when evaluating model performance metrics, a confusion matrix is ​​used to calculate precision and recall. The precision calculation formula is: The recall rate is calculated as follows: ;

[0041] Where TP is the number of true instances, TN is the number of true negative instances, FP is the number of false positive instances, and FN is the number of false negative instances.

[0042] Preferably, the method is characterized in that, in step S500, when adjusting the hyperparameters of the intelligent monitoring model, a Bayesian optimization strategy based on Gaussian processes is used to find the optimal parameter combination by constructing a probabilistic mapping relationship between the hyperparameters and the model performance indicators, including the following steps:

[0043] Step S510: Define the hyperparameter space ,in, Let i be the i-th hyperparameter;

[0044] Step S520: Use the Expected Improvement Criterion (EI) as the acquisition function, calculated as follows: ,in, This is the current optimal combination of hyperparameters. For hyperparameters Corresponding model performance metrics;

[0045] Step 530: In each iteration, select the option that makes The model is trained and its performance is evaluated using the largest combination of hyperparameters, and the probability mapping is updated until the preset number of iterations or the performance index converges.

[0046] Compared with existing technologies, the above technical solution has the following advantages:

[0047] 1. Existing intelligent monitoring systems often suffer from inaccurate status judgments and delays under complex operating conditions (such as equipment start-up and shutdown, and drastic load changes) due to complex data coupling relationships and limited feature extraction capabilities. This invention addresses the problem of insufficient feature extraction in traditional methods by using a "deep belief network algorithm" (feature extraction module) to mine key feature vectors (such as implicit coupling relationships of equipment parameters) from preprocessed data. Simultaneously, it combines an "intelligent monitoring model constructed using reinforcement learning algorithm" to dynamically adapt to changes in operating conditions through real-time interactive learning with the environment, thereby improving the real-time performance and accuracy of status prediction.

[0048] 2. Existing systems mostly use a fixed sampling rate, and the time synchronization accuracy of multi-sensor data is low (often >10ms), leading to data loss during drastic changes in operating conditions or data redundancy during stable periods, affecting subsequent analysis. In this embodiment, a "high-precision clock module" is used to align the timestamps of data collected by different sensors (synchronization error <1ms), solving the time synchronization problem; adaptive sampling rate adjustment calculation formula... According to differences in working conditions Dynamically adjust the sampling rate (increase the sampling rate when the operating conditions change drastically, and decrease it when the conditions are stable) to balance data integrity and redundancy;

[0049] 3. Existing systems handle noise from industrial environments, such as electromagnetic interference and sensor fluctuations, in a simple manner (e.g., mean filtering), which easily leads to data distortion and affects model training performance. This invention addresses this issue by employing a "median filtering algorithm based on a sliding window" (…). By replacing the original value with the median value of the data within the sliding window, extreme noise is effectively filtered out while preserving data trend characteristics.

[0050] 4. Traditional feature extraction methods (such as PCA) struggle to capture the nonlinear and high-dimensional coupling relationships in complex industrial data, leading to the loss of key features. The deep belief network in this invention addresses this by "minimizing inter-layer reconstruction error" (…). Adjusting network parameters forces the network to learn the essential characteristics of the data, rather than surface noise, thereby improving its ability to uncover implicit relationships.

[0051] 5. Existing models are trained in a multidimensional static manner, which makes it difficult to adapt to dynamic changes in working conditions (such as equipment aging and load fluctuations), resulting in a decrease in accuracy after long-term use;

[0052] 6. Existing alarm mechanisms are mostly based on fixed thresholds, failing to consider prediction bias and trend changes. This leads to false alarms (e.g., alarms triggered by normal fluctuations) or missed alarms (e.g., slow anomalies not reaching the threshold) when operating conditions fluctuate. This invention employs a "dynamic threshold + trend analysis" strategy with a dynamic threshold formula (…). Based on historical prediction deviations The threshold is adaptively adjusted; at the same time, the judgment is combined with the parameter change trend of z consecutive cycles to avoid the limitations of a single threshold.

[0053] 7. Traditional data annotation relies heavily on manual labor (e.g., expert annotation of the entire dataset), which is costly and inefficient; purely automatic annotation is prone to errors. In this invention, "semi-supervised active annotation based on uncertainty quantification" uses a predictive entropy calculation formula... High-uncertainty data (entropy value > threshold) is screened and labeled by experts, while low-uncertainty data is automatically labeled. The model is iteratively optimized to balance cost and quality.

[0054] 8. Existing hyperparameter tuning methods mostly employ grid search or random search, requiring numerous trials, resulting in low efficiency and difficulty in finding the global optimum. This invention utilizes "Bayesian optimization based on Gaussian processes," employing the expected improvement criterion (…). Construct a probabilistic mapping between hyperparameters and performance, prioritize the search for potential optimal regions, reduce the number of trials, and quickly converge to the optimal hyperparameter combination. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the intelligent monitoring system for complex working conditions and its reinforcement learning optimization method, as described in this invention.

[0056] Figure 2 This is a schematic diagram illustrating the steps of the optimization method of the intelligent monitoring system for complex working conditions and its reinforcement learning optimization method according to the present invention. Detailed Implementation

[0057] The advantages of the present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments.

[0058] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0059] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0060] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0061] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0062] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0063] In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the convenience of the description of the invention and have no specific meaning in themselves. Therefore, "module" and "part" can be used interchangeably.

[0064] See Figure 1 As shown in this embodiment, a detailed description of an intelligent monitoring system for complex operating conditions will be provided. This intelligent monitoring system includes a data acquisition module, a data preprocessing module, a feature extraction module, an intelligent monitoring model construction module, and an alarm module. These five modules are connected sequentially to form a complete workflow, enabling the monitoring of the operating status of industrial equipment under complex conditions. The data acquisition module is responsible for collecting raw data. After processing by the data preprocessing module, the data is transmitted to the feature extraction module to extract key features. The intelligent monitoring model construction module analyzes and predicts based on these features, and finally, the alarm module issues corresponding alarm signals based on the prediction results.

[0065] Data acquisition module:

[0066] Its function is to collect various operating data of industrial equipment under complex working conditions, providing raw materials for subsequent processing by the intelligent monitoring system.

[0067] Data types acquired include: sensor data such as temperature sensors (thermocouples, RTDs, capable of measuring the temperature of various parts of industrial equipment, with a measurement range covering -200℃ to 1800℃ and an accuracy of ±0.1℃), pressure sensors (such as piezoelectric and strain gauge types, used to measure fluid pressure in pipelines, internal pressure of industrial equipment, etc., with a measurement range from 0 to tens of MPa), vibration sensors (such as accelerometers, monitoring the vibration frequency and amplitude of industrial equipment during operation to determine whether there are problems such as wear or loosening), flow sensors (electromagnetic flow meters, vortex flow meters, measuring the flow rate of fluid in pipelines), process parameter data, such as stirring speed, reaction time, and raw material ratio parameters in chemical production; cutting speed, feed rate, and depth of cut in machining; and environmental parameter data, such as the operating environment of industrial equipment, including temperature (typically -40℃ to 85℃), humidity (0 to 100%RH), air pressure (close to standard atmospheric pressure, approximately 101 kPa), and dust concentration.

[0068] Data Acquisition Method: Various sensors and parameter detection devices are connected to the data acquisition module via wired (including but not limited to Ethernet, RS485 bus) or wireless (such as LoRa, NB-IoT) communication methods to achieve real-time data transmission. The acquisition frequency can be adjusted according to the importance of the industrial equipment and changes in operating conditions. For critical industrial equipment and situations with drastic changes in operating conditions, the acquisition frequency can be set to several times per second or higher. For general equipment and stable operating conditions, the acquisition frequency can be appropriately reduced to decrease the data transmission and processing burden.

[0069] Data preprocessing module:

[0070] The data preprocessing module is connected to the data acquisition module to preprocess the acquired data to improve its quality and obtain standard data.

[0071] The specific preprocessing includes cleaning, noise reduction, and normalization, each of which performs the following functions.

[0072] For data cleaning, the operational data collected by the data acquisition module may contain discrepancies (such as sudden increases or decreases in data due to sensor malfunctions) and missing values ​​(such as data lost during data transmission). For noisy data, a median filtering algorithm based on a sliding window is used for removal; its calculation formula is as follows: ,in, The filtered value. The initial value is n, where n is the half-width of the sliding window. For example, in some embodiments, the half-width of the sliding window is set to n=3, for each data point. Take the three data points before and after the first three data points, and calculate the median of these data points as the filtered data value. This effectively removes noise. For missing values, methods such as linear interpolation and polynomial interpolation can be used to fill in the gaps based on the data's trend. For example, if the temperature sensor data at a certain moment is accurate, it can be supplemented by linear estimation based on the temperature values ​​at previous and subsequent moments.

[0073] For noise reduction, the original operating data may contain noise due to factors such as electromagnetic interference and mechanical vibration in the field environment. In this case, noise reduction can also be performed based on the median filtering algorithm of the sliding window provided in the above embodiment.

[0074] For normalization, different types of operational data have different dimensions and orders of magnitude. For example, temperature is measured in °C, and pressure is measured in MPa. Direct analysis can affect the accuracy of the results. Normalization can transform the data into the interval [0,1] or [-1,1]. A commonly used method is min-max normalization, and its formula is... , where x is the original running data, and min(x) and max(x) are the minimum and maximum values ​​in the original running data, respectively. For example, converting temperature data from 0~100℃ to the range of 0-1 facilitates subsequent feature extraction and model calculation.

[0075] Feature extraction module:

[0076] The feature extraction module is connected to the data preprocessing module and uses a deep belief network algorithm to analyze the standard data, thereby extracting key feature vectors that can reflect the operating status of industrial equipment.

[0077] Deep belief networks (DBNs) are deep probabilistic generative models composed of multiple stacked restricted Boltzmann machines (RBMs). They are trained through two phases: unsupervised pre-training and supervised fine-tuning. During feature extraction, the network parameters are adjusted by minimizing the inter-layer reconstruction error, which is formulated as follows: ,in, Let M be the reconstruction error of the l-th layer, and M be the number of neurons in the l-th layer. This represents the actual output value of the j-th neuron in the l-th layer. Let be the reconstructed output value of the j-th neuron in layer l. During the unsupervised pre-training phase, each RBM layer is trained sequentially, using the output of the previous layer as the input of the next layer. The implicit features of the data are extracted by learning the probability distribution of the data. In the supervised fine-tuning phase, labeled data is used to adjust the entire network, improving the accuracy of feature extraction.

[0078] For example, in some specific embodiments, for preprocessed multidimensional data such as temperature, pressure, and vibration, deep belief networks use multi-layer nonlinear transformations to map these data into a low-dimensional feature space. The resulting key feature vectors may contain stability indicators of equipment operation, fault precursor features, etc. For instance, in motor operation monitoring, the extracted feature vectors may reflect key information such as the degree of motor speed fluctuation and current harmonic content.

[0079] Intelligent monitoring model construction module:

[0080] The intelligent monitoring model construction module uses reinforcement learning algorithm to build an intelligent monitoring model. It takes the key feature vectors obtained by the feature extraction module as input, and through interactive learning with the environment, outputs the prediction results of the equipment's operating status.

[0081] The reinforcement learning algorithm employs a deep Q-network algorithm. In reinforcement learning, the agent (i.e., the monitoring model) performs actions within the environment (i.e., the device's operating state), receiving rewards or penalties based on environmental feedback. Through continuous learning, it optimizes its action strategy to maximize cumulative rewards. The Q-value update formula for the deep Q-network algorithm is as follows: ,in, In the state Next collection action Q value, For learning rate, In the state Next collection action The reward obtained later As a discount factor, For the next state, The next action is as follows. In this intelligent monitoring system, the state of the agent is represented by a key feature vector. Actions include judging the operating status of the equipment (such as normal, slightly abnormal, or severely abnormal). The reward mechanism is set according to the degree of matching between the prediction result and the actual state. A positive reward is given for accurate prediction, and a negative reward is given for incorrect prediction.

[0082] Deep Q-networks combine the advantages of deep learning and Q-learning, approximating the Q-value function through a neural network to handle complex state spaces. For example, by inputting key feature vectors into a deep neural network, the network outputs the Q-values ​​of different actions, selects the action with the largest Q-value as the current prediction result, and updates the network parameters based on environmental feedback to continuously optimize the predictive ability of the intelligent monitoring model.

[0083] The intelligent monitoring model can predict the operating status of equipment in the future by learning from historical data and analyzing real-time data, such as "the equipment temperature will exceed the warning value within the next 10 minutes" or "the equipment may experience abnormal vibration after 30 minutes of stable operation".

[0084] Alarm module:

[0085] It connects to the intelligent monitoring model building module and determines whether the equipment is in an abnormal state based on the prediction results output by the intelligent monitoring model. If it is in an abnormal state, it issues an alarm signal.

[0086] In this embodiment, the determination of whether an abnormal state is being assessed is not based on a fixed threshold, but rather on a method combining dynamically adjusted thresholds based on prediction bias with trend analysis. The dynamic threshold calculation formula is as follows: ,in, The dynamic threshold of parameter p at the current time. Let p be the initial baseline threshold. This is the deviation adjustment factor (range: 0.3~0.8). For parameter p in the past Prediction bias at time ( , Let p be the actual value of parameter p at time q. (where p is the predicted value at time p, and w is the sliding window length). When the value of parameter p in the prediction result exceeds the dynamic threshold... If the trend of this parameter value meets the preset abnormal conditions within z consecutive sampling periods, the device is determined to be in an abnormal state.

[0087] Alarm methods include, but are not limited to, audible and visual alarms (such as emitting a buzzer or illuminating a warning light), information push (such as sending alarm information to staff's mobile APP or the computer terminal of the monitoring center, including abnormal parameters, prediction results, etc.), and linkage control (such as automatically triggering a shutdown device when serious equipment abnormalities occur to prevent the accident from escalating).

[0088] It should be noted that the intelligent monitoring system for complex operating conditions provided in this embodiment also includes a high-precision clock module. This high-precision clock module is connected to the data acquisition module to provide a time synchronization reference for the data acquisition module. When the data acquisition module collects operating data through different types of sensors, the high-precision clock module timestamps the operating data collected by different types of sensors, and the synchronization error is <1ms. The data acquisition module adaptively adjusts the sampling rate according to the dynamic changes in the operating conditions of the industrial equipment. The formula for calculating the adaptively adjusted sampling rate is: ,in, The current sampling rate, The sampling rate at the previous time step. This is an adjustment factor (with a value range of 0.1 to 0.5). The difference between the current operating condition and the operating condition at the previous moment (obtained by calculating the weighted sum of the change rates of key parameters, with a value range of 0 to 1).

[0089] The high-precision clock module functions to provide a time synchronization reference for the data acquisition module, enabling timestamp alignment of data from multiple sensors. More specifically, this high-precision clock module employs a dual-source timing scheme combining BeiDou satellite timing and a local temperature-controlled crystal oscillator. BeiDou satellite timing provides a time accuracy of ±50ns, and when the satellite signal is blocked or interfered with, it automatically switches to the local temperature-controlled crystal oscillator, whose daily frequency drift rate is ≤5× This ensures that even in the absence of satellite signal, the time deviation can still be controlled within 1ms within 30 minutes.

[0090] This high-precision clock module uses the IEEE 1588 PTP precise time protocol to synchronize time with each sensor node. In an industrial Ethernet environment, the synchronization message transmission delay can be controlled within 200ns, ultimately achieving a synchronization error of less than 1ms for multi-sensor data.

[0091] In industrial equipment operation monitoring, multiple sensors are often required to collect data from different angles, such as temperature, pressure, and vibration sensors working simultaneously. Taking a thermal power generating unit as an example, dozens of sensors, including boiler outlet temperature sensors, turbine speed sensors, and furnace pressure sensors, are distributed in different locations. Only by ensuring that the data collected by these sensors are synchronized can the correlation between parameters such as "whether a sudden increase in temperature leads to a synchronous increase in pressure" be accurately analyzed. If the timing is not synchronized, it may be misjudged as "pressure rises first, then temperature rises" instead of "temperature rises first, then pressure rises," leading to incorrect analysis of the causal relationship of the fault. A synchronization error of less than 1ms can minimize the impact of such errors.

[0092] The adaptive sampling rate adjustment of the data acquisition module dynamically adjusts the data sampling rate based on changes in the equipment's operating conditions, optimizing the data volume while ensuring no critical information is lost. More specifically, the formula for adaptive sampling rate adjustment is as follows: , The current sampling rate, The sampling rate at the previous time step. This is an adjustment factor (with a value range of 0.1 to 0.5). The difference between the current operating condition and the operating condition at the previous moment (obtained by calculating the weighted sum of the change rates of key parameters, with a value range of 0 to 1).

[0093] for The calculation method is obtained by calculating the weighted sum of the rates of change of key parameters, and the formula is as follows: Where n is the number of key parameters (e.g., selecting 5 to 8 core parameters such as temperature, pressure, and speed). The weight of the i-th parameter (set according to the degree of influence of the parameter on the equipment state, with a total of 1, such as a speed weight of 0.3, a pressure weight of 0.25, etc.) This represents the value of the i-th parameter at the current time. The value of the i-th parameter at the previous time step. This is the rated reference value for the parameter (e.g., rated speed 3000 r / min). When the operating conditions of the equipment change drastically (e.g., the stirring speed of a chemical reactor suddenly increases from 50 r / min to 300 r / min), the rate of change of the key parameter increases. The value is close to 1. According to the formula, the current sampling rate The sampling rate will be significantly increased from the previous time step, allowing for more intensive data collection and the capture of detailed information during dramatic changes in operating conditions (such as vibration peaks during sudden speed changes). When the equipment operates stably (e.g., the steam turbine is stable at its rated speed of 3000 r / min), the rate of change of key parameters approaches zero. The value is close to 0, and the current sampling rate It will maintain the level of the previous moment or decrease slightly to reduce data redundancy (e.g., during stable operation, there is no need to collect repetitive, stable data at high frequency). The sampling rate has upper and lower limit protection, with the lower limit... (e.g., Hz, to ensure no loss of basic data), upper limit (e.g., 2000Hz, to avoid hardware overload due to excessively high sampling rates), when the calculated When the value is outside the range, the boundary value will be automatically taken.

[0094] To facilitate understanding, this embodiment will be described in more detail through specific implementation methods. For example, in the monitoring panel of a high-speed rail traction converter, the data acquisition module is responsible for collecting data from various sensors such as voltage, current, temperature, and vibration. The key parameter selected is current (…). ), rotational speed ( ),temperature( ),vibration( The reference values ​​are 1000A, 300km / h, 80℃, and 0.1g. When the traction converter switches from constant speed (200km / h) operation to accelerated operation, the previous parameter values ​​are: current 800A, speed 200km / h, temperature 60℃, and vibration 0.05g; the current parameter values ​​are: current 1000A, speed 220km / h, temperature 65℃, and vibration 0.08g. Calculate the rate of change of each parameter: current... Rotation speed ,temperature ,vibration ;but: ,like Hz, The current sampling rate is: By increasing the sampling rate, details of current fluctuations during acceleration are captured. Once the operation stabilizes (e.g., at an average speed of 300 km / h), data from five consecutive time points are collected. All values ​​are less than 0.05, at which point the sampling rate gradually decreases. =500Hz.

[0095] In the data preprocessing module described in the above embodiments, when preprocessing the running data to obtain standard data, the preprocessing includes cleaning. This embodiment will again describe the cleaning process in more detail. Specifically, when cleaning the running data, the data preprocessing module uses a sliding window median filtering algorithm to remove noisy data. The cleaning calculation formula is as follows: ,in, The filtered value. is the initial value, and n is the half width of the sliding window.

[0096] The principle of window-based median filtering is that it's a non-linear filtering method where the median of all data within a sliding window is used to replace the data value at the center of the window. Compared to linear filtering methods like mean filtering, median filtering is more effective at removing impulse noise (such as sudden abnormal jumps in sensor values) and salt-and-pepper noise, while also better preserving signal edge information and avoiding data distortion. For this cleaning calculation formula... ,in, The filtered value. Let n be the initial value, where n is half the width of the sliding window, and the total width of the window is 2n+1. For example, when n=3, the total width of the window is 7, which is taken as... arrive The median of these 7 data points is used as... .

[0097] The selection of the sliding window's half-width 'n' directly affects the filtering effect and needs to be set reasonably based on the data's frequency of change, considering the noise characteristics. If 'n' is too small, the window contains too little data, resulting in poor denoising; if 'n' is too large, although it can effectively remove noise, it may smooth out useful abrupt changes in the data. In practical applications, it is usually adjusted according to the data's sampling frequency and noise intensity, generally ranging from 1 to 5. For slowly changing data (such as ambient temperature), 'n' can be increased appropriately; for rapidly changing data (such as vibration signals), 'n' needs to be decreased to retain key instantaneous change information.

[0098] The specific processing flow during cleaning is as follows: First, determine the half-width n of the sliding window; slide the window from left to right on the original motion data sequence, sliding one data point at a time; for each window position, calculate the median of all data within the window; replace the original data value at the center position of the window with the calculated median to obtain the filtered data.

[0099] For edge data processing, when the window slides to the edge of the data sequence, such as... (The first data in the sequence), if n=2, the window needs to contain... , , , , ,but and This does not exist. In this case, data completion is usually used, such as mirroring the edge data (using...). Alternative , Alternative Alternatively, the median of existing data can be taken to ensure that edge data can also be effectively filtered.

[0100] To facilitate understanding, the content described in this embodiment will be further described in detail through specific embodiments. For example, in the current data acquisition of a steel production equipment, due to strong electromagnetic interference on site, the acquired current data is often mixed with noise, manifested as sudden increases or decreases in some data points. Assume the original current data sequence (unit: A) is: [50,52,180,53,51,54,49,55,200,53], where 180 and 200 are obvious noise points. Select n=2, that is, the total window width is 5; process... When =180, the data in the window is =52、 =180、 =53、 , =54, and arranging these data points gives [51,52,53,54,180], with a median of 53. Therefore... =53, successfully removed the noise point at 180. (During processing...) At that time, the data in the window is =49、 =55、 =200、 Due to insufficient edge data, supplementation is needed. (Mirrored expansion), the window data is [49,53,55,200,53], after sorting it is [49,53,53,55,200], the median is 53, therefore =53, removing the noise point 200. The filtered sequence is: [50,52,53,53,51,54,49,55,53,53], which effectively eliminates noise while preserving the overall trend of the original running data.

[0101] In this embodiment, the labeled data will be further analyzed using a feature extractor to obtain key feature vectors that reflect the operating status of industrial equipment. The inter-layer reconstruction error formula during feature extraction using a deep belief network algorithm is as follows: ,in, Let M be the reconstruction error of the l-th layer, and M be the number of neurons in the l-th layer. This represents the actual output value of the j-th neuron in the l-th layer. This represents the reconstructed output value of the j-th neuron in layer l. In this embodiment, the adjustment logic of the calculation formula is as follows: by calculating the reconstruction error of each layer, the connection weights and other parameters between that layer and its adjacent layers are adjusted according to the magnitude of the error. When the reconstruction error is large, it indicates that the network's processing effect on the data of that layer is poor, requiring significant parameter adjustments; when the reconstruction error is small, the network parameters are in a better state, enabling accurate feature extraction.

[0102] Reconstruct output value It is obtained through the backpropagation process of a deep belief network. Taking the l-th layer as an example, it uses the actual output value of that layer. Based on the current network parameters, the input information of the previous layer is derived in reverse. Then, based on this input information of the previous layer, the reconstructed output value of the l-th layer is calculated in the forward direction, so as to compare the difference between the actual output and the reconstructed output.

[0103] The number of neurons, M, in the l-th layer affects the calculation of reconstruction error. The choice of M needs to be determined based on the complexity of the data and the requirements of feature extraction. If M is too small, it may not be able to fully capture the data features, resulting in a large reconstruction error; if M is too large, it will increase the complexity and computational cost of the network, and may also lead to overfitting.

[0104] The parameter tuning process involves initializing the parameters of each layer of the deep belief network, such as connection weights and biases; then performing forward propagation on the network to obtain the actual output values ​​of each neuron. Perform backpropagation to calculate the reconstructed output values ​​of neurons in each layer. Using optimization algorithms such as gradient descent, based on Adjust the parameters of this layer to reduce the reconstruction error; repeat the above steps until the reconstruction error of each layer reaches the preset minimum value or no longer decreases significantly.

[0105] To facilitate understanding, this embodiment will be described in more detail through specific examples. For instance, in the feature extraction of operational data from electronic manufacturing equipment, a certain intermediate layer (let's call it layer 2) of a deep belief network has M=50 neurons. After forward propagation of the network, the actual output value of each neuron in this layer is obtained. (j=1,2,...,50); the reconstructed output values ​​are obtained through backpropagation and forward computation. ; Calculate the reconstruction error of this layer ;like Larger, for example, the initial calculation yielded =0.8 indicates that the parameter settings for this layer are unreasonable, and the connection weights and other parameters between the second layer and the first and third layers need to be adjusted; after several parameter adjustments, When the value is reduced to 0.1, the layer can process the data better, and the extracted features such as the temperature change rate of electronic components and current stability are more accurate, providing high-quality input for the subsequent prediction of the intelligent monitoring model.

[0106] In the above embodiments, the intelligent monitoring model construction module was described. Its core function is to use the deep Q-network algorithm to implement reinforcement learning, and to optimize the intelligent monitoring model by continuously updating the Q value, so that the intelligent monitoring model can accurately output the prediction result of the operating status based on the key feature vector of the device operation. In this embodiment, the deep Q-network algorithm used in the reinforcement learning algorithm will be described in detail, including the following contents.

[0107] The Deep Q-Network algorithm is a reinforcement learning algorithm that combines deep learning and Q-learning. The deep learning component (usually a deep neural network) is used to fit the Q-value function, capable of handling high-dimensional state spaces; Q-learning is a value-based reinforcement learning method that guides the agent to make optimal decisions by learning the value (Q-value) of actions. The Q-value update formula... ,in, In the state Next collection action Q value, For learning rate, In the state Next collection action The reward obtained later As a discount factor, For the next state, This is the next action. Regarding the meaning of the Q-value in this calculation formula, the Q-value represents the expected cumulative reward of taking a certain action in a specific state. The higher the Q-value, the more likely the action is to bring a better result in the current state. For example, in a state of abnormal equipment temperature, if the Q-value of taking the "emergency stop" action is higher than that of "continue observation," the model will tend to choose "emergency stop." The key feature vectors obtained by the feature extraction module in the above embodiments constitute the core indicators of equipment operation, such as temperature change rate and pressure fluctuation amplitude; actions This refers to the decisions that the intelligent monitoring model can make regarding the device's status, such as "determine normal operation," "issue a minor anomaly warning," or "issue a serious anomaly alarm." The reward function is designed to reward... The settings directly affect the learning direction of the intelligent monitoring model. For example, when the intelligent monitoring model predicts the device status and the actual status is consistent, a positive reward (e.g., +10) is given; when the intelligent monitoring model misses a fault (the actual fault is predicted to be normal), a smaller negative reward (e.g., -20) is given. Regarding the update process of this calculation formula, in the formula... The target Q value represents the state. Take action below Then, the sum of the immediate reward that can be obtained and the discounted reward of the best future action. This represents the temporal difference error, reflecting the difference between the current Q-value and the target Q-value. The learning rate... By controlling the magnitude of the adjustment towards the target Q value during each update, and through continuous iteration, the Q value gradually approaches the optimal value.

[0108] For ease of understanding, this embodiment will be described in detail through a specific example. For instance, in the training of an intelligent monitoring model for mining equipment, the equipment status at a certain moment... The key feature vectors show "bearing temperature slightly high (55℃), vibration frequency slightly exceeding the standard". The intelligent monitoring model adopts dynamic... "Issued a minor anomaly warning"; after actual operation, no serious malfunctions occurred, and an award was given. =+5; Next state In the process, the intelligent monitoring model calculates the Q-value of the optimal action. Discount factor =0.9, learning rate =0.1; Initial =6, according to the calculation formula 6 + 0.1 × (5 + 0.9 × 8 - 6) = 6.62, completing one Q-value update; after multiple iterations, the Q-value of the intelligent monitoring model taking the action of "issuing a minor anomaly warning" in this state gradually stabilizes, indicating that the model has learned the optimal decision in this state.

[0109] In the above embodiments, the function of the alarm module was described. In this embodiment, it will be described in more detail again. Specifically, when the alarm module determines whether the industrial equipment is in an abnormal state, it makes the judgment based on a combination of dynamic adjustment threshold for prediction deviation and trend analysis. The calculation formula for the dynamic adjustment threshold is: ,in, The dynamic threshold of parameter p at the current time. Let p be the initial baseline threshold. This is the deviation adjustment factor (range: 0.3~0.8). For parameter p in the past Prediction bias at time ( , Let p be the actual value of parameter p at time q. (where p is the predicted value of parameter, and w is the sliding window length); when the value of parameter p in the prediction result exceeds the dynamic threshold... Furthermore, if the trend of parameter value changes meets the preset abnormal conditions within z consecutive sampling periods, the industrial equipment is determined to be in an abnormal state. In this embodiment, since equipment parameters often undergo dynamic fluctuations under complex operating conditions, fixed thresholds are prone to false alarms or missed alarms. The content of this embodiment is designed to adapt to the changing patterns of parameters, thereby improving the accuracy and timeliness of alarms.

[0110] For the calculation formula of the dynamically adjusted threshold In this calculation formula The dynamic threshold for the current time p; The initial baseline threshold for parameter p (set according to the equipment manual or historical normal data). This is the deviation adjustment factor (ranging from 0.3 to 0.8, adjusted according to the importance of the parameter, with a larger value for key parameters). The average prediction bias of parameter p over the past w time points is calculated using the following formula: , Let p be the actual value of parameter p at time q. is the predicted value at time p, and w is the sliding window length, ranging from 5 to 20.

[0111] The logic for this alarm module to determine whether industrial equipment is in an abnormal state is as follows: when the value of parameter p in the prediction result exceeds the dynamic threshold... If the trend of this parameter value meets the preset abnormal conditions within z consecutive sampling periods, the device is judged to be in an abnormal state. Here, z is the number of sampling periods (set according to the parameter change rate, such as z=3~5 for rapidly changing parameters and z=10~15 for slowly changing parameters).

[0112] For adaptive adjustment of dynamic thresholds, when the parameter prediction deviation is large ( When the dynamic threshold increases, It will automatically loosen to avoid false alarms caused by prediction errors; when the prediction deviation decreases ( When the threshold is reduced, the dynamic threshold It will tighten, increasing the rigor of the judgment. For example, in some embodiments, a certain pressure parameter... =10MPa, =0.5, w=10, if the past 10 moments =0.8MPa, then the current =10×(1+0.5×0.8)=14MPa, which is more adaptable to prediction bias compared to a fixed threshold.

[0113] Preset abnormal conditions for trend analysis include parameter change rates continuously exceeding thresholds (e.g., temperature rises by more than 2°C per second), parameter fluctuation amplitudes continuously increasing (e.g., vibration amplitude standard deviation continuously rising), and parameters continuously deviating from the normal range. It should be noted that trend analysis can also incorporate the physical characteristics of the parameters. For example, an abnormal trend for rotational speed parameters might be "continuous acceleration exceeding the safe slope," while for liquid level parameters it might be "continuous decline exceeding the design value."

[0114] To facilitate understanding, this embodiment will be described in more detail through a specific example. Taking the monitoring of the generator speed parameter p of a wind turbine generator set as an example, for each parameter... =1800r / min (baseline threshold). =0.5 (deviation adjustment coefficient), w=10 (sliding window length), z=5 (number of consecutive sampling periods), the preset abnormal condition is "speed change rate > 0.05 r / (min・s)"; average prediction deviation is calculated based on the actual values ​​of the past 10 time points. Compared with the predicted value If the sum of the absolute values ​​of the deviations is 200 r / min, then =200 / 10=20r / min; Dynamic threshold calculation =1800×(1+0.5×20 / 1800)=1800×(1+0.0056)=1810r / min.

[0115] In the anomaly detection process, if the intelligent monitoring model predicts that the current rotational speed is 1820 r / min (exceeding...), Within five consecutive sampling periods (1 second per period), the rotational speed change rates were 0.06, 0.07, 0.08, 0.07, and 0.09 r / (min·s), respectively, all meeting the preset abnormal conditions; the alarm module determined that the generator was in an abnormal state and issued a level two alarm (such as pushing early warning information to the operation and maintenance terminal).

[0116] See Figure 2 As shown, the present invention also provides a reinforcement learning optimization method for an intelligent monitoring system for complex working conditions, comprising the following steps:

[0117] Step S100: Collect a large amount of historical operating data from industrial equipment under complex conditions. This data covers sensor data, process parameter data, and environmental parameter data under various operating states (normal, minor fault, severe fault, etc.). Data sources include the equipment's historical database and records stored in real-time monitoring systems. The completeness and diversity of the data must be ensured to cover different complex operating conditions. The collected data is then labeled, i.e., each data point is tagged with the equipment's operating status. The types of tags are determined based on the possible states the equipment may experience, such as "normal operation," "abnormal temperature," "excessive pressure," and "abnormal vibration." The labeling process can combine human experience with automated tools. For complex status judgments, review by domain experts is required to ensure the accuracy of the labeling.

[0118] Step S200: Divide the labeled data into a training set and a test set according to a certain ratio. The training set is used to train the model, enabling it to learn the patterns and features in the data; the test set is used to evaluate the performance of the trained model and test its generalization ability to new data. This division includes, but is not limited to, a ratio of 70% training set and 30% test set.

[0119] Step S300: Set the initial weights, biases, and other parameters of the deep Q-network, as well as the initial values ​​of hyperparameters such as the learning rate and discount factor; input the samples from the training set into the model, and the intelligent monitoring model outputs the prediction results. Calculate the Q-value according to the Q-value update formula in the aforementioned embodiment and adjust the parameters of the intelligent monitoring model. Through continuous iteration, the intelligent monitoring model gradually learns the mapping relationship between input features and device operating status, improving the accuracy of prediction. It should be noted that during the training process, the loss function value of the intelligent monitoring model on the training set is calculated periodically. When the loss function value tends to be stable and small, it indicates that the intelligent monitoring model has reached a good training state.

[0120] Step S400: Test the trained intelligent monitoring model using a test set to evaluate its performance metrics, such as accuracy and recall. Input the samples from the test set into the trained model to obtain the prediction results. Calculate accuracy and recall based on the prediction results and the actual labels of the test set samples. Accuracy reflects the proportion of correct predictions made by the intelligent monitoring model, while recall reflects the model's ability to identify actual abnormal states. These metrics provide a comprehensive understanding of the intelligent monitoring model's performance.

[0121] Step S500: Based on the test results, adjust the hyperparameters of the intelligent monitoring model, such as the learning rate and discount factor, and retrain and test until the intelligent monitoring model's performance reaches its optimal level. If the intelligent monitoring model's performance is poor, such as low accuracy or insufficient recall, the cause needs to be analyzed and the hyperparameters adjusted. For example, when the intelligent monitoring model overfits, the learning rate can be reduced or the regularization term can be increased; when the intelligent monitoring model converges too slowly, the learning rate can be appropriately increased. After adjusting the hyperparameters, retrain the model using the training set and test the model's performance using the test set. Through multiple iterative adjustments, the performance of the intelligent monitoring model will gradually improve until it reaches the preset optimal standard.

[0122] For ease of understanding, this embodiment will be described in more detail through a specific example. In the optimization of an intelligent monitoring model for a ship's power equipment, the optimization steps include the following:

[0123] Step S100: Collect historical operating data of the ship's power equipment under different navigation conditions (such as full load, no load, different speeds, etc.), and label the equipment with status tags such as normal, minor fault, and serious fault. A total of 10,000 data points were collected.

[0124] Step S200: Divide the data into a training set (7000 records) and a test set (3000 records) in a 7:3 ratio.

[0125] Step S300: Train the intelligent monitoring model using the deep Q-network algorithm on the training set, with an initial learning rate of 0.1 and a discount factor of 0.9. During training, the loss function is calculated every 100 iterations, and the loss function tends to stabilize after 5000 iterations.

[0126] Step S400: Evaluate the model performance using the test set, obtaining an accuracy of 85% and a recall of 80%.

[0127] Step S500: Since the accuracy and recall did not meet expectations (target accuracy 90%, recall 85%), the learning rate was adjusted to 0.05, and the model was retrained. After three adjustments, the model achieved an accuracy of 92% and a recall of 88%, meeting the performance requirements.

[0128] During the annotation process in step S100 above, the semi-supervised active annotation strategy based on uncertainty quantification specifically includes the following steps:

[0129] Step S110: Train an initial classification model based on the labeled data, and calculate the prediction entropy value of the unlabeled data. The calculation formula is as follows: ,in, Let be the predicted entropy value of data d, and K be the total number of categories. The model predicts the probability that data d belongs to class k. Specifically, in this step, an initial classification model is first trained using a small amount of labeled data (e.g., 10%-20% of the total data). This initial model serves to make preliminary predictions about the class of unlabeled data, providing a foundation for subsequent uncertainty quantification. The initial model can employ a simple classification algorithm (such as support vector machines or decision trees) or a lightweight neural network. Then, using a large amount of unlabeled data, the prediction entropy value of the initial classification model is calculated to quantify the uncertainty of the initial classification model regarding the data class. The calculation formula is as follows: In the formula denoted as the predicted entropy value of data d. The higher the entropy value, the less uncertain the initial classification model is about the category of data d. K is the total number of categories (e.g., if the equipment operating status is divided into normal, minor fault, and severe fault, then K=3). Let be the probability that the initial classification model predicts data d as belonging to category k, and the sum of the probabilities of all categories is 1.

[0130] Step S200: Obtain labeled data with predicted entropy values ​​greater than a preset threshold, which are then labeled by domain experts. Specifically, a preset threshold for predicted entropy values ​​is set (e.g., 1.0, adjusted based on the number of categories and the performance of the initial classification model), and unlabeled data with entropy values ​​greater than this threshold is selected. This type of data, due to the high uncertainty in the initial classification model's predictions, contains more potential information and is considered "high-value data." It is manually labeled by domain experts to ensure labeling accuracy.

[0131] Step S300: Add the newly labeled data to the training set and iteratively update the initial classification model until the average prediction entropy of the unlabeled data is lower than the convergence threshold. Specifically, add the newly labeled data to the training set, retrain the initial classification model (at this point, the model accuracy has improved), and then recalculate the prediction entropy of the remaining unlabeled data using the updated model, repeating the "filter-label-update" process. When the average prediction entropy of the unlabeled data is lower than the preset convergence threshold (e.g., 0.5), it indicates that the model has a relatively certain understanding of the category of most unlabeled data, and no further manual labeling is needed; the labeling process ends.

[0132] Regarding the meaning of the prediction entropy value mentioned in the above embodiments, entropy is an indicator in information theory that measures uncertainty. When the model's prediction probability for data d is concentrated in a certain category (such as...), the entropy value is used to measure uncertainty. (Other categories have probabilities close to 0). A value close to 0 indicates that the initial classification model makes a clear judgment; when the probability distribution is dispersed (e.g., ...), the probability distribution is not clear. , , ),at this time A large value indicates that the initial classification model's judgment is ambiguous, and the data needs to be manually labeled.

[0133] Regarding the principles for setting the above thresholds, the prediction threshold (for filtering high-value data) needs to be adjusted based on the K value. The larger the K value, the higher the maximum possible entropy value (e.g., when K=3, the maximum entropy value is...). The threshold is usually set to 70% to 80% of the maximum possible entropy value. For the convergence threshold (termination annotation): it needs to be set according to the application scenario's requirements for annotation accuracy. The higher the requirements, the lower the threshold (e.g., for monitoring critical equipment, it can be set to 0.3).

[0134] The initial labeled data should be representative and cover the main operating states of the equipment (such as normal operation and various typical faults) to avoid bias in the initial classification model due to uneven data distribution. This includes, but is not limited to, using a stratified sampling method to randomly select a small number of data points from each state category as initial labeled samples.

[0135] For ease of understanding, this embodiment will be described in more detail again through a specific example. For example, in the labeling of aircraft engine operating data (category K=3; normal, minor fault, serious fault).

[0136] Step S100: Select 100 labeled data points (30 normal, 30 with minor faults, and 40 with severe faults) to train the initial classification model; for 1000 unlabeled data points, the probability of the initial classification model predicting a certain data point d is: , , ,but .

[0137] Step S200: If the preset threshold is 1.0, the data is because... Selected items are labeled as "minor faults" by experts.

[0138] Step S300: Add the newly labeled data to the training set, update the model, and recalculate the entropy value. After 5 iterations, the average predicted entropy value of the unlabeled data decreased from the initial 1.0 to 0.4 (below the convergence threshold of 0.5), and the labeling ended. At this point, only 200 manually labeled data entries were accumulated (far fewer than the 1000 fully labeled data entries), but the labeling quality met the model training requirements.

[0139] In this embodiment, step S400 in the aforementioned embodiment will be described in detail. When evaluating the model performance indicators in step S400, accuracy and recall are calculated using a confusion matrix to quantitatively evaluate the performance of the intelligent monitoring model. Specifically, the confusion matrix is ​​a matrix used to show the relationship between the prediction results and actual results of the intelligent monitoring model. It divides the predictions of the initial classification model into four types: true instances (TP), true negative instances (TN), false positive instances (FP), and false negative instances (FN). Among them, true instances (TP) are when the intelligent monitoring model accurately predicts the abnormal state of the device (actually abnormal, and predicted as abnormal), true negative instances (TN) are when the intelligent monitoring model correctly predicts the normal state of the device (actually normal, and predicted as normal), false positive instances (FP) are when the intelligent monitoring model incorrectly predicts the normal state as an abnormal state (actually normal, and predicted as abnormal), and false negative instances (FN) are when the intelligent monitoring model incorrectly predicts the abnormal state as a normal state (actually abnormal, and predicted as normal).

[0140] Accuracy refers to the proportion of samples correctly predicted by the intelligent monitoring model out of the total number of samples. Its calculation formula is: Accuracy reflects the overall predictive accuracy of the intelligent monitoring model and is suitable for evaluating its comprehensive performance on balanced datasets. Recall (also known as sensitivity or recall) refers to the proportion of anomalous state samples correctly predicted by the intelligent monitoring model to the actual number of anomalous state samples, calculated as follows: Recall rate focuses on measuring the ability of intelligent monitoring models to identify abnormal states, which is especially important in industrial monitoring because missed anomalies (false negatives) can lead to serious equipment failures and safety accidents.

[0141] The confusion matrix mentioned in this embodiment is constructed by first inputting samples from the test set into a trained intelligent monitoring model to obtain a prediction result (normal or abnormal) for each sample; then comparing the actual state of the sample with the prediction result, and counting the number of TP, TN, FP, and FN to construct the confusion matrix. For example, in 100 test samples, there are actually 20 abnormal samples and 80 normal samples. The intelligent monitoring model predicts 18 abnormal samples (15 of which are correct and 3 are incorrectly predicted as normal) and 82 normal samples (78 of which are correct and 4 are incorrectly predicted as abnormal). Therefore, TP=15, TN=78, FP=4, and FN=5.

[0142] Regarding precision and recall, the closer the precision value is to 1, the higher the overall proportion of correct predictions made by the intelligent monitoring model. However, it cannot solely reflect the model's ability to identify abnormal states. For example, when the proportion of abnormal samples in the dataset is extremely low, the intelligent monitoring model may achieve a high precision even if all predictions are normal, but the recall will be very low. Conversely, the closer the recall value is to 1, the more actual abnormal samples the intelligent monitoring model can identify, resulting in a lower false negative rate. In industrial scenarios, it is typically necessary to prioritize a high recall rate for the intelligent monitoring model to minimize the chance of missing anomalies.

[0143] It's important to note that precision and recall often have a trade-off. Increasing recall may lead to more false positives and lower precision; conversely, increasing precision may lead to more false negatives and lower recall. In practical applications, a balanced weighting of these two metrics needs to be determined based on the specific monitoring requirements and the importance of the equipment, in order to optimize the performance of the intelligent monitoring model. For example, for critical equipment, a higher recall rate might be prioritized, even at the cost of some precision; for general equipment, a higher precision rate can be pursued while maintaining a certain level of recall.

[0144] To facilitate understanding, this embodiment will be described in detail through a specific example. For instance, in the intelligent monitoring model test of an automobile production line, 1000 equipment operating status samples are selected as the test set, including 100 actual abnormal samples and 900 normal samples. After inputting the samples into the intelligent monitoring model, the following results are obtained: True Positive Cases (TP): 85 (the intelligent monitoring model correctly predicted 85 abnormal samples), True Negative Cases (TN): 880 (the intelligent monitoring model correctly predicted 880 normal samples), False Positive Cases (FP): 20 (the intelligent monitoring model incorrectly predicted 20 normal samples as abnormal), False Negative Cases (FN): 15 (the intelligent monitoring model incorrectly predicted 15 abnormal samples as normal). Regarding the calculation of accuracy, This means the overall accuracy of the intelligent monitoring model is 96.5%, indicating that it can make correct predictions in most cases. Regarding the calculation of recall rate... The intelligent monitoring model achieved a recall rate of 85%, meaning that out of 100 actual abnormal samples, it successfully identified 85 and missed 15. This indicates that while the intelligent monitoring model has high accuracy, its recall rate has room for improvement. Since equipment malfunctions on automotive production lines can lead to production interruptions and product quality issues, the consequences of missed detections are severe. Therefore, the intelligent monitoring model needs optimization, such as adjusting hyperparameters and increasing training data for abnormal samples, to improve recall and reduce the false negative rate.

[0145] In this embodiment, the process of adjusting the hyperparameters of the intelligent monitoring model in step S500 will be described in detail. When adjusting the hyperparameters of the intelligent monitoring model, the optimal parameter combination is found by constructing a probabilistic mapping relationship between the hyperparameters and the model performance indicators based on the Bayesian optimization strategy of Gaussian process. Specifically, the following steps are included.

[0146] Step S510: First, define the hyperparameter space. ,in, Let be the i-th hyperparameter. For the definition of the hyperparameter space, the hyperparameter space... The determination of hyperparameters is not arbitrary; it requires consideration of the characteristics of the intelligent monitoring model and the needs of the actual industrial scenario. For example, for an intelligent monitoring model built using a deep Q-network algorithm, hyperparameters may include the learning rate (affecting the magnitude of model parameter updates), the discount factor (reflecting the weight of future rewards in the current decision), the number of layers in the neural network, and the number of neurons per layer (determining the model's fitting ability), etc. Each hyperparameter has its reasonable range of values; for example, the learning rate is usually between 0.001 and 0.1, and the discount factor is generally between 0.8 and 0.99. These ranges are determined based on extensive experiments and industry experience, avoiding excessively extreme hyperparameter values ​​that could lead to model training failure or poor performance. After determining the hyperparameter space, a Gaussian process is used to construct a probabilistic mapping relationship between hyperparameters and model performance indicators. A Gaussian process is a nonparametric model based on a multivariate normal distribution. It can probabilistically describe the distribution of performance indicators within the entire hyperparameter space using existing combinations of hyperparameters and their corresponding performance indicator data. Specifically, for any two combinations of hyperparameters in the hyperparameter space... , Their corresponding performance indicators and The joint distribution of follows a multivariate normal distribution, and its mean and covariance are determined by the mean function and covariance function of the Gaussian process.

[0147] Step S520: Use the Expected Improvement Criterion (EI) as the acquisition function, and its calculation formula is: ,in, This is the current optimal combination of hyperparameters. For hyperparameters The corresponding model performance metrics are explained by breaking down the calculation formula. This represents the expectation operation. This indicates a new combination of hyperparameters. performance indicators Equivalent to the current optimal combination of hyperparameters performance indicators The expected improvement value is the estimated improvement amount, taking into account both the potential performance improvement from the new hyperparameter combination and the probability of such improvement occurring. For example, when the predictive performance of a hyperparameter combination is significantly higher than the current best and the prediction uncertainty is low, its expected improvement value will be large; conversely, when the predictive performance is lower than the current best, the expected improvement value will be 0. By calculating the expected improvement value for each possible hyperparameter combination, we can prioritize the combinations most likely to bring performance improvements for subsequent model training and evaluation.

[0148] Step S530: In each iteration, select the hyperparameter combination with the largest expected improvement value and use this combination to train the intelligent monitoring model. The training process strictly follows the model training flow, including data input, forward propagation to calculate prediction results, loss function calculation, and backpropagation to update model parameters. After training, the model's performance metrics (such as accuracy, recall, F1 score, etc.) are evaluated using a test set. The newly obtained hyperparameter combination and its corresponding performance metric data are added to the existing dataset, and the Gaussian process model is updated again, i.e., the mean function and covariance function are recalculated to more accurately describe the probabilistic mapping relationship between hyperparameters and performance metrics. The above steps of calculating the acquisition function and selecting hyperparameters are repeated until a stopping condition is met. The stopping condition can be a preset maximum number of iterations (e.g., 50 times), or the improvement in model performance metrics in multiple consecutive iterations being less than a preset threshold (e.g., 0.005), or the maximum expected improvement value being less than a certain threshold, indicating that continuing the search may not yield significant performance improvements.

[0149] To facilitate understanding, this embodiment will be explained in detail through a specific example. The hyperparameter space of the intelligent monitoring model includes the learning rate. 0.01-0.1 Discount Factor (0.8-0.99), Number of neurons in the hidden layer of the neural network Taking (32-128, step size 16) as an example, the iterative optimization process is explained in detail. In the initial stage, the following hyperparameter combinations are randomly selected: (0.05, 0.9, 64), (0.03, 0.85, 32), and (0.08, 0.95, 128). The model is trained and its performance evaluated for these three combinations, yielding accuracies of 0.82, 0.78, and 0.85, respectively. At this point, the current optimal hyperparameter combination... The optimal performance index is (0.08, 0.95, 128). Based on these three sets of data, a Gaussian process model was constructed, and the expected improvement value of each combination in the hyperparameter space was calculated. It was assumed that the hyperparameter combination (0.06, 0.93, 96) yielded the largest expected improvement value. Using this combination to train the model, an accuracy of 0.88 was obtained, which is better than the current best result; therefore, the model was updated. The values ​​are (0.06, 0.93, 96). This set of data was added to the dataset, and the Gaussian process model was updated. In the second iteration, the expected improvement for each hyperparameter combination was recalculated, and the combination (0.07, 0.94, 80) was found to have the largest expected improvement. The accuracy of the trained model was 0.89, and it was updated again. The model was iterated further. In the 10th iteration, the hyperparameter combination (0.065, 0.935, 90) achieved an accuracy of 0.90. The 11th iteration used the combination (0.063, 0.932, 88), which achieved an accuracy of 0.902 after training, with an improvement of 0.002. The 12th iteration used the combination (0.064, 0.933, 89), achieving an accuracy of 0.903, with an improvement of 0.001. Since the performance improvement in two consecutive iterations was less than the preset threshold of 0.005, the stopping condition was met, and the iteration was stopped. The final optimal hyperparameter combination was (0.064, 0.933, 89), with a corresponding model accuracy of 0.903, indicating that the model performance was optimal.

[0150] It should be noted that the embodiments of the present invention have better implementability and are not intended to limit the present invention in any way. Any person skilled in the art may use the above-disclosed technical content to change or modify it into equivalent effective embodiments. However, any modifications or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. An intelligent monitoring system for complex working conditions, characterized in that, include: The data acquisition module is used to collect operating data of industrial equipment under complex working conditions; A data preprocessing module, connected to the data acquisition module, is used to preprocess the running data acquired by the data acquisition module to obtain standard data; The feature extraction module is connected to the data preprocessing module and uses a deep belief network algorithm to extract features from the standard data to obtain key feature vectors that can reflect the operating status of the industrial equipment. The intelligent monitoring model construction module uses a reinforcement learning algorithm to construct an intelligent monitoring model. The intelligent monitoring model takes the key feature vector as input and outputs the prediction result of the operating status of the industrial equipment after interactive learning. An alarm module, connected to the intelligent monitoring model construction module, is used to determine whether the industrial equipment is in an abnormal state based on the prediction results, and to issue an alarm signal if the industrial equipment is in an abnormal state.

2. The intelligent monitoring system for complex working conditions according to claim 1, characterized in that, It also includes a high-precision clock module connected to the data acquisition module. The high-precision clock module provides a time synchronization reference for the data acquisition module. When the data acquisition module collects the operating data through different types of sensors, the high-precision clock module timestamps the operating data collected by different types of sensors, and the synchronization error is <1ms. The data acquisition module adaptively adjusts the sampling rate based on the dynamic changes in the operating conditions of the industrial equipment. The formula for calculating the adaptively adjusted sampling rate is as follows: in, The current sampling rate, The sampling rate at the previous time step. This is an adjustment factor (with a value range of 0.1 to 0.5). The difference between the current operating condition and the operating condition at the previous moment (obtained by calculating the weighted sum of the change rates of key parameters, with a value range of 0 to 1).

3. The intelligent monitoring system for complex working conditions according to claim 1, characterized in that, The preprocessing of the data preprocessing module includes cleaning, denoising, and normalizing the running data to obtain the standard data; wherein, When cleaning the running data, the data preprocessing module uses a sliding window median filtering algorithm to remove noisy data. The cleaning calculation formula is as follows: in, The filtered value. is the initial value, and n is the half width of the sliding window.

4. The intelligent monitoring system for complex working conditions according to claim 1, characterized in that, During feature extraction, the deep belief network algorithm adjusts network parameters by minimizing inter-layer reconstruction error. The formula for calculating the inter-layer reconstruction error is: in, Let M be the reconstruction error of the l-th layer, and M be the number of neurons in the l-th layer. This represents the actual output value of the j-th neuron in the l-th layer. This is the reconstructed output value of the j-th neuron in layer l.

5. The intelligent monitoring system for complex working conditions according to claim 1, characterized in that, The reinforcement learning algorithm employs a deep Q-network algorithm, and the Q-value update formula for the deep Q-network algorithm is as follows: in, In the state Next collection action Q value, For learning rate, In the state Next collection action The reward received later As a discount factor, For the next state, For the next action.

6. The intelligent monitoring system for complex working conditions according to claim 1, characterized in that, When determining whether the industrial equipment is in an abnormal state, the alarm module makes a judgment based on a combination of dynamic adjustment threshold for prediction deviation and trend analysis. The formula for calculating the dynamic adjustment threshold is as follows: in, The dynamic threshold of parameter p at the current time. Let p be the initial baseline threshold. This is the deviation adjustment factor (range: 0.3~0.8). For parameter p in the past Prediction bias at time ( , Let p be the actual value of parameter p at time q. (where p is the predicted value of parameter, and w is the sliding window length); when the value of parameter p in the prediction result exceeds the dynamic threshold... If the trend of parameter value change meets the preset abnormal conditions within z consecutive sampling periods, the industrial equipment is determined to be in an abnormal state.

7. A reinforcement learning optimization method for an intelligent monitoring system under complex working conditions, characterized in that, Includes the following steps: Step S100: Collect historical operating data of industrial equipment under complex working conditions and label it, labeling the industrial equipment with tags under different operating states; Step S200: Divide the labeled running data into a training set and a test set; Step S300: Train the intelligent monitoring model based on the training set, and update the parameters of the intelligent monitoring model using a deep Q-network algorithm during the training process; Step S400: Test the intelligent monitoring model with updated parameters based on the test set to evaluate the performance indicators of the intelligent monitoring model with updated parameters and obtain the test results; Step S500: Based on the test results, adjust the hyperparameters of the intelligent monitoring model, and train and test again until the performance of the intelligent monitoring model reaches the set standard.

8. The reinforcement learning optimization method for an intelligent monitoring system for complex working conditions according to claim 7, characterized in that, In step S100, the semi-supervised active annotation strategy based on uncertainty quantification includes the following steps: Step S110: Train an initial classification model based on the labeled data, and calculate the prediction entropy value of the unlabeled data. The calculation formula is: in, Let be the predicted entropy value of data d, and K be the total number of categories. Predict the probability that data d belongs to category k for the model; Step S120: Obtain labeled data with predicted entropy values ​​greater than a preset threshold, and have them labeled by domain experts; Step S130: Add the newly labeled data to the training set and iteratively update the initial classification model until the average prediction entropy of the unlabeled data is lower than the convergence threshold.

9. The reinforcement learning optimization method for an intelligent monitoring system for complex working conditions according to claim 7, characterized in that, In step S400, when evaluating model performance metrics, a confusion matrix is ​​used to calculate precision and recall. The precision calculation formula is as follows: The recall rate is calculated as follows: ; Where TP is the number of true instances, TN is the number of true negative instances, FP is the number of false positive instances, and FN is the number of false negative instances.

10. The reinforcement learning optimization method for an intelligent monitoring system for complex working conditions according to claim 7, characterized in that, In step S500, when adjusting the hyperparameters of the intelligent monitoring model, a Bayesian optimization strategy based on Gaussian processes is used to find the optimal parameter combination by constructing a probabilistic mapping relationship between hyperparameters and model performance indicators, including the following steps: Step S510: Define the hyperparameter space ,in, Let i be the i-th hyperparameter; Step S520: Use the Expected Improvement Criterion (EI) as the acquisition function, calculated as follows: ,in, This is the current optimal combination of hyperparameters. For hyperparameters Corresponding model performance metrics; Step S530: In each iteration, select the option that makes The model is trained and its performance is evaluated using the largest combination of hyperparameters, and the probability mapping is updated until the preset number of iterations or the performance index converges.

Images