Industrial data prediction method based on AI machine learning

Abstract

The invention relates to the technical field of industrial data prediction, and discloses an industrial data prediction method based on AI machine learning. According to the method, the stability and integrity of data input are judged after hierarchical preprocessing; when the data input is unstable or incomplete, the fluctuation amplitude of the data acquisition frequency and the accumulated duration of transmission delay are analyzed in a collaborative manner, the reliability level of a data acquisition link is evaluated and divided into a high level, a middle level and a low level, and a corresponding link optimization strategy is executed; when the grade is medium reliability, time sequence feature extraction is carried out on the preprocessed data, and the stability and periodicity of data distribution are evaluated; and finally, dynamically adjusting the input weight and the iteration frequency of the prediction model by adopting a reinforcement learning algorithm according to the link reliability level and the data distribution characteristics. According to the method, the data acquisition link state and the data distribution characteristics are dynamically adapted, so that the prediction model can more accurately respond to the dynamic data change in industrial production, and the prediction performance is improved.

Description

Technical Field

[0001] This invention relates to the field of industrial data prediction technology, specifically to an industrial data prediction method based on AI machine learning. Background Technology

[0002] In industrial production, accurate data prediction plays an irreplaceable role in ensuring smooth production processes, rational resource allocation, and early avoidance of potential risks. With the advancement of industrial intelligence, the number of sensors, controllers, and other equipment deployed in production lines has surged, generating increasingly complex data types covering multiple dimensions such as temperature, pressure, speed, and energy consumption. The real-time nature, completeness, and stability of this data directly affect the output of predictive models, and consequently, the effectiveness of production decisions.

[0003] Traditional industrial data forecasting methods often have significant limitations in the data acquisition stage. Most methods rely on a single type of sensor or fixed acquisition nodes, making it difficult to cover the multi-dimensional variables in the production environment, resulting in one-sided data collection. For example, a traditional forecasting system in an automobile manufacturing workshop only collects equipment operating temperature data, ignoring factors such as humidity and voltage fluctuations. When a sudden increase in humidity causes equipment failure, the forecasting model cannot provide early warning. Furthermore, data preprocessing often uses simple filtering or normalization operations, lacking a hierarchical processing mechanism. This makes it difficult to effectively remove noisy data and outliers, leading to prediction biases when directly fed into the model.

[0004] In terms of data transmission and link management, existing technologies generally lack a systematic assessment of the reliability of data acquisition links. During the transmission of data from the acquisition end to the processing end, problems such as fluctuating acquisition frequency and accumulated transmission delays often occur due to network fluctuations, equipment aging, etc. Traditional methods either ignore these problems and directly input unstable data into the model, or adopt a uniform link repair scheme, performing the same adjustment measures regardless of the actual link status, resulting in low link optimization efficiency.

[0005] The parameter settings of the predictive model are another key factor affecting the prediction results. Traditional models often use fixed input weights and iteration frequencies, which are not adjusted after training. However, the distribution characteristics of industrial data often change with production rhythms and environmental variations, sometimes showing a stable trend and sometimes exhibiting obvious periodic fluctuations. When the data distribution changes, or the reliability of the data acquisition link fluctuates, fixed parameters cannot adapt to the new data source characteristics, leading to a greater deviation between the prediction results and the actual situation. For example, the raw material consumption data of a chemical production line shows different periodicities during the day and night, but the traditional model uses the same iteration frequency, resulting in a significant deviation between the nighttime predicted value and the actual consumption, affecting the accuracy of raw material allocation.

[0006] In existing technologies, the correlation between the status of the data acquisition link and the characteristics of data distribution is often overlooked. When the link reliability is high, the data integrity is good, and the model may not require high-frequency iteration; however, when the link reliability is moderate, the data may fluctuate, requiring more flexible parameter adjustments. Traditional methods fail to establish this correlation mechanism, making the model insufficiently adaptable to complex and ever-changing industrial environments, further limiting the improvement of prediction accuracy. The combination of these problems makes industrial data prediction difficult to fully utilize in practical applications and unable to meet the demands of modern industrial production for precise decision-making. Summary of the Invention

[0007] The purpose of this invention is to provide an industrial data prediction method based on AI machine learning to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides an industrial data prediction method based on AI machine learning, the method comprising:

[0009] S1: During the parameter configuration of the initial prediction model, dynamic data in the industrial production environment is acquired in real time through a multi-dimensional data acquisition unit. After the dynamic data is preprocessed in layers, the stability and integrity of the data input are judged.

[0010] S2: When data input is unstable or incomplete, perform a joint analysis of the fluctuation range of data acquisition frequency and the cumulative duration of transmission delay to assess the reliability level of the data acquisition link;

[0011] S3: Based on the evaluation results, the reliability level of the data acquisition link is divided into high reliability, medium reliability and low reliability, and the corresponding link optimization strategy is executed.

[0012] S4: When the reliability level of the data acquisition link is medium reliability, perform time series feature extraction on the preprocessed industrial data to evaluate the stationarity and periodicity of the data distribution.

[0013] S5: Based on the reliability level of the data acquisition link and the stability and periodicity of the data distribution, the input weights and iteration frequency of the prediction model are dynamically adjusted using reinforcement learning algorithms.

[0014] Preferably, in S1, a multi-dimensional data acquisition unit is deployed in the industrial production environment, including vibration sensors, current sensors, and humidity sensors. Real-time data is collected from the sensors at a preset cycle and recorded as a raw dataset. The raw dataset includes the acquisition time, vibration amplitude, current intensity, and ambient humidity. The data integrity is labeled, and the labeled data is divided into a training sample set and a validation sample set for building and validating an adaptive filtering model. An adaptive filtering algorithm is used to initialize the filtering coefficients. The adaptive filtering model is trained using the training sample set, and the filtering effect of the model is verified using the validation sample set. The multi-dimensional data acquisition unit collects vibration amplitude, current intensity, and ambient humidity data in the industrial production environment in real time to form dynamic input data. The dynamic input data is input into the trained adaptive filtering model, and the model outputs the data noise coefficient based on the current data.

[0015] The noise coefficient of the data output by the adaptive filtering model is compared with the preset noise threshold. If the noise coefficient is greater than or equal to the preset noise threshold, the data input is unstable or incomplete, and a data cleaning instruction is triggered. If the noise coefficient is less than the preset noise threshold, the data input is stable and complete, and the subsequent process continues.

[0016] Preferably, in step S2, a frequency anomaly index is generated based on the fluctuation amplitude of the data acquisition frequency. The method for obtaining the frequency anomaly index is as follows:

[0017] The clock module deployed in the data acquisition unit records the time interval of each data acquisition in real time and calculates the frequency deviation value for each acquisition cycle. Specifically, the calculation method is: the difference between the standard time interval and the actual time interval for two consecutive acquisition cycles; the smoothed frequency deviation value is calculated using a sliding window averaging method, where the smoothing process involves superimposing the current window's frequency deviation value with the smoothing result of the previous window according to a preset weight; the obtained smoothed frequency deviation value is compared with a pre-set standard frequency deviation range to calculate the frequency anomaly index, where the standard frequency deviation range is the interval between the mean of historical frequency deviations under normal operating conditions and three times the standard deviation; when the smoothed frequency deviation value exceeds the upper limit of the standard frequency deviation range, the frequency anomaly index is calculated as: the smoothed frequency deviation value minus the upper limit of the standard frequency deviation range; when the smoothed frequency deviation value is lower than the lower limit of the standard frequency deviation range, the frequency anomaly index is calculated as: the lower limit of the standard frequency deviation range minus the smoothed frequency deviation value; where the frequency anomaly index is a quantitative indicator measuring the degree of fluctuation in the data acquisition frequency.

[0018] Preferably, a delay impact index is generated based on the cumulative duration of transmission delay within a fixed time period. The method for obtaining the delay impact index is as follows:

[0019] The system acquires the data transmission delay time for each transmission within a specified time period and records it as a time series dataset. It then uses the Empirical Mode Decomposition (EMD) method to perform intrinsic mode function (IMF) decomposition on the time series dataset, obtaining multiple IMF components of different frequencies. The system calculates the energy percentage of each IMF component, which is the ratio of the variance of a single IMF component to the sum of the variances of all components. Finally, it calculates the delay impact index by combining the energy percentages of each component with their corresponding delay times. The specific calculation method is the sum of the products of the energy percentages of each component and their corresponding delay times.

[0020] Preferably, the frequency anomaly index and the delay impact index are converted into a second feature vector, which is then used as the input to the deep belief network model. The deep belief network model uses the prediction of the reliability level label of the data acquisition link for each set of second feature vectors as the training objective, and minimizes the classification error of all reliability level labels as the optimization direction. The deep belief network model is trained until the classification error converges to a preset range and then training stops. The reliability level of the data acquisition link is determined based on the model output.

[0021] Preferably, in S3, based on the evaluation results, the reliability level of the data acquisition link is divided into high reliability, medium reliability, and low reliability, specifically as follows:

[0022] The reliability level assessment value of the acquired data acquisition link is compared with a preset dual threshold, which includes a first reliability threshold and a second reliability threshold, and the first reliability threshold is less than the second reliability threshold. The reliability level assessment value is then compared with the first reliability threshold and the second reliability threshold respectively.

[0023] If the reliability level assessment value is greater than the second reliability threshold, it is classified as high reliability and the current data acquisition and transmission parameters are maintained; if the reliability level assessment value is greater than or equal to the first reliability threshold and less than or equal to the second reliability threshold, it is classified as medium reliability and the data acquisition frequency and transmission bandwidth are adjusted to ensure a balance between data real-time performance and integrity; if the reliability level assessment value is less than the first reliability threshold, it is classified as low reliability, the backup data link is immediately activated, and a link failure warning is issued.

[0024] Preferably, in step S4, a data distribution characteristic index is generated based on the preprocessed industrial data to evaluate the stationarity and periodicity of the data distribution. The method for obtaining the data distribution characteristic index is as follows:

[0025] Preprocessed industrial data within a fixed time period is acquired, and the data is detrended to eliminate the influence of long-term trends. The window size and smoothing coefficient of the local weighted regression model are determined, and the local weighted regression model is used to fit the detrended data to obtain the local trend curve of the data. The deviation between the actual data value and the local trend curve is calculated, and the absolute values ​​of the deviation values ​​are accumulated as the data distribution characteristic index. The data distribution characteristic index is a comprehensive indicator that measures the degree of deviation of the data from the local trend. The smaller the value, the more stable the data distribution.

[0026] Preferably, the obtained data distribution characteristic index is compared with a preset data stability threshold. If the data distribution characteristic index is greater than or equal to the preset data stability threshold, the data distribution stability is low, and a data stability abnormal signal is generated. If the data distribution characteristic index is less than the preset data stability threshold, the data distribution stability is high, and a data stability normal signal is generated.

[0027] Preferably, in S5, the input weights and iteration frequency of the prediction model are dynamically adjusted using a reinforcement learning algorithm based on the reliability level of the data acquisition link and the stability and periodicity of the data distribution.

[0028] The reliability level assessment value of the data acquisition link and the data distribution characteristic index are used as the state input of the reinforcement learning agent, and the input weight adjustment amount and iteration frequency adjustment amount of the prediction model are used as the action output of the reinforcement learning agent.

[0029] The reward value of the current action is calculated through a preset reward function, with the reduction in prediction error and the model convergence speed as the core indicators.

[0030] The agent's policy network and value network are updated using a deep deterministic policy gradient algorithm to optimize the action selection policy.

[0031] Based on the optimal action output by the agent, the input weights and iteration frequency of the prediction model are dynamically adjusted to achieve real-time optimization of prediction accuracy.

[0032] Preferably, in S5, when dynamically adjusting the input weights and iteration frequency of the prediction model using a reinforcement learning algorithm, multi-objective optimization of the prediction model is also included. Specifically, prediction accuracy, model running efficiency, and resource utilization are used as input parameters of the multi-objective optimization function, weight coefficients of each parameter are set, and the multi-objective optimization function is solved using a non-dominated sorting genetic algorithm to obtain the Pareto optimal solution set. Based on the real-time requirements of the industrial scenario, a target solution is selected from the Pareto optimal solution set, and the model parameters corresponding to the target solution are used as the initial policy of the reinforcement learning agent. Then, the policy is optimized using a deep deterministic policy gradient algorithm.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] This AI-based machine learning-based industrial data prediction method acquires dynamic data from the industrial production environment in real time through a multi-dimensional data acquisition unit and performs hierarchical preprocessing. This allows it to capture various variables in the production process from multiple perspectives. Furthermore, the hierarchical processing refines the data filtering process, reducing the interference of noisy data and outliers on subsequent analysis. In the data input stability and integrity assessment stage, it can promptly identify potential issues such as missing or fluctuating data during acquisition, preventing invalid data from directly entering the model calculation stage and providing a higher-quality input foundation for the model.

[0035] When data input is unstable or incomplete, a synergistic analysis of the fluctuation range of data acquisition frequency and the cumulative duration of transmission delay can comprehensively assess the actual status of the data acquisition link. This synergistic analysis does not view a single indicator in isolation, but combines frequency fluctuations with cumulative delay to more accurately reflect the true reliability of the link and avoid misjudging the link status based on a single indicator.

[0036] Based on the assessment results, the reliability levels of the data acquisition links are divided into high, medium, and low, and corresponding optimization strategies are implemented to achieve targeted and flexible link management. At the high reliability level, the existing link status can be maintained to reduce unnecessary adjustments; at the medium reliability level, link stability can be improved through specific optimization measures; at the low reliability level, an enhanced repair mechanism is activated to quickly improve link quality and avoid resource waste or incomplete repairs caused by using a uniform solution.

[0037] When the link reliability level is medium, extracting time series features from preprocessed industrial data and assessing the stationarity and periodicity of the data distribution can reveal deeper underlying patterns. Data at the medium reliability level often exhibits some fluctuations; analyzing its stationarity and periodicity at this level can capture the characteristics of data changes over time, providing a more specific basis for model adjustments.

[0038] Based on link reliability levels and data distribution characteristics, reinforcement learning algorithms are used to dynamically adjust the input weights and iteration frequency of the prediction model, enabling the model to possess autonomous adaptability. When link status changes or data distribution characteristics alter, the model can optimize parameters in real time through reinforcement learning, adapting to new data source characteristics without human intervention. For example, when link reliability decreases, the model can increase the iteration frequency to quickly track data changes; when data exhibits significant periodicity, the model can adjust input weights to highlight the impact of periodic characteristics. This dynamic adaptation mechanism allows the prediction model to maintain good predictive performance in complex and ever-changing industrial environments, reducing prediction bias caused by environmental changes. Attached Figure Description

[0039] Figure 1 This is a schematic diagram illustrating the working principle of the AI ​​machine learning-based industrial data prediction method described in this invention.

[0040] Figure 2 A flowchart for generating the frequency anomaly index;

[0041] Figure 3 A flowchart illustrating the delayed impact on index generation;

[0042] Figure 4 This is a flowchart for generating and evaluating data distribution characteristic indices. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Please see Figure 1 This invention provides an industrial data prediction method based on AI machine learning, the method comprising:

[0045] S1: During the parameter configuration of the initial prediction model, dynamic data from the industrial production environment is acquired in real time through a multi-dimensional data acquisition unit. After hierarchical preprocessing of the dynamic data, the stability and completeness of the data input are assessed. The multi-dimensional data acquisition unit can collect information from different dimensions of the industrial production environment, ensuring the comprehensiveness of the acquired data. Hierarchical preprocessing involves classifying and processing the data according to its type, source, etc., to prepare for subsequent assessments of data stability and completeness.

[0046] S2: When data input is unstable or incomplete, a collaborative analysis is performed on the fluctuation range of the data acquisition frequency and the cumulative duration of transmission delay to assess the reliability level of the data acquisition link. Fluctuations in data acquisition frequency and transmission delay both affect data quality; collaborative analysis of these two factors provides a more comprehensive understanding of the data acquisition link's status, thereby accurately assessing its reliability.

[0047] S3: Based on the evaluation results, the reliability level of the data acquisition link is divided into high reliability, medium reliability, and low reliability, and corresponding link optimization strategies are implemented. Different reliability levels correspond to different link problems, and targeted optimization strategies can effectively improve the link condition and ensure normal data acquisition and transmission.

[0048] S4: When the reliability level of the data acquisition link is medium, perform time series feature extraction on the preprocessed industrial data to assess the stationarity and periodicity of the data distribution. Time series feature extraction can uncover the patterns of data changes over time, while assessing the stationarity and periodicity of the data distribution helps to make more accurate data predictions in the future.

[0049] S5: Based on the reliability level of the data acquisition link and the stability and periodicity of the data distribution, a reinforcement learning algorithm is used to dynamically adjust the input weights and iteration frequency of the prediction model. This dynamic adjustment enables the prediction model to adapt to different link conditions and data distribution characteristics, improving the accuracy and adaptability of the prediction.

[0050] Example 1: See Figure 2 In S1, a multi-dimensional data acquisition unit is deployed in the industrial production environment. This unit integrates vibration sensors, current sensors, and humidity sensors. These sensors monitor different physical quantities in the industrial production environment. The vibration sensor captures changes in vibration amplitude during equipment operation, the current sensor records fluctuations in current intensity in the circuit in real time, and the humidity sensor continuously collects the humidity of the production environment. Real-time data is synchronously collected from these sensors according to a pre-set time period, and the collected data is summarized to form a raw dataset. The raw dataset contains information such as the time of each acquisition, the vibration amplitude value recorded by the vibration sensor, the current intensity data fed back by the current sensor, and the environmental humidity parameter measured by the humidity sensor. The data integrity in the raw dataset is labeled, that is, each data point is labeled according to whether it contains complete information on the four items: acquisition time, vibration amplitude, current intensity, and environmental humidity, thereby distinguishing complete data from incomplete data. After labeling, these data are divided into a training sample set and a validation sample set. The training sample set is used to build an adaptive filtering model, and the validation sample set is used to test the actual performance of the model.

[0051] When using an adaptive filtering algorithm, the filter coefficients are first initialized, with initial values ​​set based on historical data characteristics of the industrial production environment. Then, the training sample set is input into the adaptive filtering model. The model dynamically adjusts the filter coefficients by continuously learning the data patterns in the training samples. During training, the model repeatedly compares the output results with the actual sample values ​​and continuously optimizes its internal parameters based on these differences. After training, a validation sample set is introduced to verify the model's filtering effect. By analyzing the output of the data in the validation sample set after filtering, it is determined whether the model can effectively filter out noise components in the data. Once the model's filtering effect meets expectations, it can be put into practical application. Vibration amplitude, current intensity, and ambient humidity data from the industrial production environment are collected in real time by a multi-dimensional data acquisition unit. This real-time data constitutes the dynamic input data, which is then input into the trained adaptive filtering model. The model calculates and outputs the corresponding data noise coefficient based on the characteristics of the current input data.

[0052] The noise coefficient of the data output by the adaptive filtering model is compared with a preset noise threshold, which is determined based on the basic data quality requirements in industrial production. If the noise coefficient is greater than or equal to the preset noise threshold, it indicates that the collected data contains a lot of noise, and the data input is unstable or incomplete. In this case, a data cleaning command will be automatically triggered. After the data cleaning command is initiated, the system will process the current dynamic input data, including removing obviously abnormal data points and filling in missing data fields, to reduce the impact of noise on subsequent processes. If the noise coefficient is less than the preset noise threshold, it indicates that the currently collected data has little noise, and the data input is stable and complete. In this case, no data cleaning operation is required, and the subsequent processing flow can continue directly.

[0053] In S2, when data input is unstable or incomplete, it is necessary to analyze the fluctuation range of the data acquisition frequency to generate a frequency anomaly index. To achieve this analysis, a clock module is deployed inside each data acquisition unit. This module can accurately record the specific time of each data acquisition and obtain the actual acquisition cycle length by calculating the time interval between two consecutive data acquisitions. The frequency deviation value of each acquisition cycle is calculated by the difference between the standard time interval and the actual time interval of two consecutive acquisition cycles. The standard time interval is the preset acquisition cycle length, and the actual time interval is calculated from the time recorded by the clock module. A sliding window averaging method is used to smooth the frequency deviation value. During processing, the frequency deviation value in the current window is superimposed with the smoothed result of the previous window according to preset weights to obtain the smoothed frequency deviation value. The allocation of preset weights can be set according to the historical characteristics of data fluctuations to balance the influence of recent data and historical data on the smoothing result.

[0054] A pre-defined standard frequency deviation range is established, calculated based on historical frequency deviation data under normal operating conditions. Specifically, this range comprises the interval between the mean of historical frequency deviations and three times the standard deviation. The smoothed frequency deviation value is compared to this standard frequency deviation range to calculate the frequency anomaly index. When the smoothed frequency deviation value exceeds the upper limit of the standard frequency deviation range, the frequency anomaly index is the smoothed frequency deviation value minus the upper limit of the standard frequency deviation range; conversely, when the smoothed frequency deviation value is below the lower limit of the standard frequency deviation range, the frequency anomaly index is the lower limit of the standard frequency deviation range minus the smoothed frequency deviation value. The frequency anomaly index calculated in this way quantifies the degree of fluctuation in the data acquisition frequency, providing an important indicator for assessing the reliability of the data acquisition link.

[0055] While analyzing the fluctuation range of data acquisition frequency, it is also necessary to conduct a collaborative analysis in conjunction with the cumulative duration of transmission delay. Transmission delay refers to the time consumed in the process of data being sent from the acquisition unit to the processing system. By recording the delay time of each data transmission, the total transmission delay within a fixed period is calculated. Combining the frequency anomaly index with the analysis results of the cumulative transmission delay, the overall reliability level of the data acquisition link is comprehensively evaluated from two dimensions: the time stability of data acquisition and transmission efficiency. This provides a basis for judgment on subsequent link optimization strategies.

[0056] Example 2: See Figure 3In S2, to address situations where data input is unstable or incomplete, it's necessary to analyze the cumulative transmission delay over a fixed period to generate a delay impact index. First, the system acquires the delay time for each data transmission within a specified timeframe, arranging these delay times chronologically to form a time-series dataset. This dataset contains detailed records of the time elapsed for each data transmission from the acquisition unit to the processing center, covering fluctuations in transmission delay at different times.

[0057] The aforementioned time series dataset was processed using the Empirical Mode Decomposition (EMD) method. This method can decompose complex time series into multiple intrinsic mode components (EMCs) with different frequency characteristics. Each EMC represents a fluctuation component within a specific frequency range of the original time series. Through this decomposition, different features contained in transmission delay, such as high-frequency instantaneous fluctuations, mid-frequency periodic variations, and low-frequency trend variations, can be separated, making the physical meaning of each component clearer. During the decomposition process, no basis functions need to be preset; the decomposition is entirely based on the characteristics of the data itself, thus better reflecting the actual fluctuation patterns of transmission delay data.

[0058] After completing the intrinsic mode decomposition, the energy percentage of each intrinsic mode component is calculated. The energy percentage is obtained by dividing the variance of a single intrinsic mode component by the sum of the variances of all intrinsic mode components. Variance reflects the dispersion of the data; the larger the variance of a particular intrinsic mode component, the greater its proportion in the overall fluctuation of transmission delay. By calculating the energy percentage, the contribution of different frequency components to the total fluctuation of transmission delay can be quantified. For example, if the energy percentage of a certain component is high, it indicates that fluctuations within that frequency range are one of the main factors causing changes in transmission delay.

[0059] The delay impact index is calculated by combining the energy proportion of each intrinsic mode component with its corresponding delay time. Specifically, the energy proportion of each component is multiplied by its corresponding delay time, and the sum of all products is the delay impact index. This calculation method can comprehensively consider the influence of different frequency components on the cumulative effect of transmission delay, reflecting both the dominant role of high-energy-proportion components and the superposition effect of low-energy-proportion components, thus more comprehensively reflecting the cumulative effect of transmission delay over a fixed duration.

[0060] The previously obtained frequency anomaly index and delay impact index are converted into a second feature vector. During this conversion, the two indices need to be standardized to eliminate the influence of differences in dimensions, bringing them to the same order of magnitude. The standardized frequency anomaly index and delay impact index, as two dimensions of the second feature vector, together constitute a feature representation that reflects the cumulative effect of data acquisition frequency fluctuations and transmission delays. This vector form facilitates subsequent input into the deep belief network model for processing.

[0061] The second feature vector is input into a deep belief network model, which is composed of multiple stacked Restricted Boltzmann Machines (RBMs) and possesses powerful feature learning and classification capabilities. The model's training process aims to obtain link reliability level labels from pre-defined data corresponding to each set of second feature vectors. By continuously adjusting the weights and bias parameters in the network, the difference between the model's output reliability level label and the pre-defined label is gradually reduced. The optimization direction of training is to minimize the classification error for all reliability level labels; that is, the prediction error is calculated using the backpropagation algorithm, and the network parameters are updated according to the error magnitude, iterating this process repeatedly.

[0062] During training, a preset range for classification error needs to be set. When the model's classification error decreases to this range and no longer changes significantly, the model is considered to have converged, and training is stopped. After training, the deep belief network model can automatically identify the feature patterns contained in the input second feature vector and quickly output the corresponding reliability level assessment result. In this way, the reliability level of the data acquisition link can be accurately determined based on the cumulative effect of fluctuations in data acquisition frequency and transmission delay, providing a basis for subsequent link optimization strategies.

[0063] In practical applications, the model can quickly calculate and output a reliability level for a new input second feature vector, without requiring manual intervention, thus achieving automated assessment of the reliability of the data acquisition chain. Furthermore, because the model is trained on a large amount of historical data, its evaluation results can adapt well to the complex changes in the data acquisition chain within industrial production environments.

[0064] Example 3: See Figure 4 In S3, based on the reliability level assessment results of the data acquisition link, it is divided into high reliability, medium reliability, and low reliability levels. This classification process relies on preset dual thresholds: a first reliability threshold and a second reliability threshold, with the first reliability threshold being less than the second reliability threshold. These two thresholds are determined based on the historical performance and actual needs of data transmission in industrial production, and can reflect the basic quality level of data transmission under different link conditions.

[0065] The reliability level assessment value is compared with the first reliability threshold and the second reliability threshold. If the reliability level assessment value is greater than the second reliability threshold, it indicates that the current state of the data acquisition link can stably transmit data, and the data loss rate, transmission latency, and other indicators are at a low level. Therefore, it is classified as a high reliability level. At this level, no adjustment is needed to the link parameters; maintaining the current data acquisition frequency and transmission bandwidth is sufficient, and the data can be transmitted to the processing system completely according to the expected rhythm.

[0066] If the reliability level assessment value is greater than or equal to the first reliability threshold and less than or equal to the second reliability threshold, it indicates that the link status is at a medium level, and it is classified as a medium reliability level. At this level, although the link has not experienced a serious failure, the stability or real-time performance of data transmission may fluctuate to some extent, requiring performance optimization by adjusting the data acquisition frequency and transmission bandwidth. During the adjustment process, the requirements of industrial production for data real-time performance and data integrity must be comprehensively considered. For example, in critical production stages, the transmission bandwidth can be appropriately increased to ensure timely data arrival, while in non-critical stages, the acquisition frequency can be appropriately reduced to decrease the data volume, thus achieving a balance between the two.

[0067] If the reliability level assessment value is less than the first reliability threshold, it indicates a significant link anomaly, potentially leading to frequent data loss or prolonged transmission delays, and is classified as a low reliability level. At this level, to prevent data acquisition interruptions from impacting subsequent processing, the system immediately activates a backup data link. The backup link uses different transmission paths and hardware than the primary link, enabling seamless takeover of data transmission tasks in the event of a primary link failure. Simultaneously, the system issues a link failure warning, which is sent to relevant monitoring terminals in text or audio-visual format, prompting maintenance personnel to inspect the primary link.

[0068] In S4, when the reliability level of the data acquisition link is medium, time series feature extraction is required on the preprocessed industrial data to assess the stationarity and periodicity of the data distribution and generate a data distribution characteristic index. First, preprocessed industrial data within a fixed time period is acquired. This data has undergone previous filtering and cleaning processes to remove some noise and outliers, thus better reflecting the true state of the industrial production process.

[0069] Detrending the acquired industrial data aims to eliminate long-term trend components. Long-term trends, which may be caused by factors such as equipment aging and production expansion, can mask short-term fluctuations and affect assessments of stability and periodicity if left untreated. Detrending is achieved by calculating the deviation between the data and the long-term trend line. Specifically, linear or polynomial fitting can be used to determine the long-term trend, which is then subtracted from the original data to obtain the detrended data.

[0070] Determine the window size and smoothing coefficient for the local weighted regression model. The window size determines the range of historical data the model references during the fitting process. An excessively large window may lead to over-smoothing and loss of local features, while an excessively small window may make the model overly sensitive to noise. The smoothing coefficient controls the smoothness of the fitted curve; a larger value results in a smoother curve, and vice versa. These two parameters need to be adjusted based on the data sampling frequency and fluctuation characteristics to adapt to the data features of different industrial scenarios.

[0071] A locally weighted regression model is used to fit the detrended data to obtain a local trend curve. When fitting each data point, the model assigns different weights to data within its surrounding window; data points closer to the data have higher weights. This allows the fitted curve to better follow the local trend of the data. This fitting method can more accurately capture the fluctuation patterns of data over a short period.

[0072] Calculate the deviation between the actual data value and the local trend curve, and sum the absolute values ​​of the deviations to obtain the data distribution characteristic index, i.e.:

[0073]

[0074] Where F represents the data distribution characteristic index, n represents the total number of data points, and x i This represents the i-th actual data value. This represents the fitted value of the local trend curve at the i-th data point. This represents the absolute value of the deviation between the i-th actual data value and the fitted value. This represents the summation of the absolute values ​​of the deviations over all data points. The data distribution characteristic index comprehensively reflects the degree to which data deviates from the local trend. The smaller the value, the higher the fit between the actual data and the local trend curve, and the more stable the data distribution; the larger the value, the more drastic the data fluctuations and the worse the stability. This index can be used to quantitatively assess the distribution characteristics of industrial data over time series, providing a basis for adjusting the parameters of subsequent prediction models.

[0075] Example 4: In S4, after extracting time-series features from the preprocessed industrial data, it is necessary to evaluate the stationarity and periodicity of the data distribution. This is specifically achieved by comparing the obtained data distribution characteristic index with a pre-set data stationarity threshold. The data stationarity threshold is determined based on the characteristic distribution of historical industrial production data. The threshold may differ in different industrial scenarios. For example, in the precision instrument manufacturing scenario, the allowable range of data fluctuation is narrow, so the threshold is set relatively low; while in the heavy machinery production scenario, the data itself fluctuates more, so the threshold is set relatively high.

[0076] If the data distribution characteristic index is greater than or equal to a pre-set data stationarity threshold, it indicates that the data fluctuations over time exceed the normal range, and the data distribution stationarity is low. In this case, the system will generate a data stationarity anomaly signal. This signal will be transmitted to the data processing center to trigger subsequent targeted processing mechanisms. For example, in a chemical production process, if the data distribution characteristic index reflecting changes in reactor temperature exceeds the threshold, the generated anomaly signal will indicate that the temperature data fluctuates significantly, possibly due to factors such as unstable stirring rate or fluctuations in raw material feed rate.

[0077] If the data distribution characteristic index is less than a pre-set data stationarity threshold, it indicates that the data fluctuations over time are within a normal range and the data distribution has high stationarity. In this case, a normal data stationarity signal is generated. This signal is also transmitted to the data processing center, indicating that the current data characteristics conform to normal production patterns and can be used for subsequent predictive analysis according to normal procedures. For example, in a car assembly line, if the data distribution characteristic index reflecting the conveyor belt speed is below the threshold, the generated normal signal indicates that the conveyor belt is operating stably and the speed fluctuations are within a reasonable range.

[0078] In S5, reinforcement learning algorithms are used to dynamically adjust the input weights and iteration frequency of the prediction model based on the reliability level of the data acquisition link and the stability and periodicity of the data distribution. By constructing an interaction mechanism between the agent and the environment, reinforcement learning enables the agent to learn the optimal strategy through continuous trial and error, making it suitable for dynamically changing scenarios in industrial data prediction.

[0079] The reliability level assessment value and data distribution characteristic index of the data acquisition link are used as state inputs for the reinforcement learning agent. The reliability level assessment value reflects the current state of the data transmission link. For example, if the sensor data transmission link of a production line is assessed as having a medium reliability level, it indicates that there is a certain risk of transmission delay or data loss. The data distribution characteristic index reflects the fluctuation characteristics of the data itself. For example, a low distribution characteristic index of current data over a certain period of time indicates that the current changes relatively steadily. These two parameters together form the basis for the agent's perception of the environment, enabling the agent to fully understand the current data acquisition quality and data characteristics.

[0080] The input weight adjustments and iteration frequency adjustments of the prediction model are used as the action outputs of the reinforcement learning agent. The input weight adjustments determine the importance of different dimensions of data in the prediction model. For example, when predicting equipment failure, the weight of vibration data may need to be increased, while the weight of environmental humidity data may be appropriately decreased. The iteration frequency adjustments determine the time interval for updating model parameters. When data fluctuates drastically, the iteration frequency can be increased to allow the model to adapt to new data features more quickly, while when data is stable, the iteration frequency can be decreased to reduce computational resource consumption.

[0081] The reward value for the current action is calculated using a pre-defined reward function, with the reduction in prediction error and the model convergence speed as the core metrics. A higher reward value is awarded when the agent's output adjustment reduces the prediction error and accelerates model convergence; conversely, a lower reward value is awarded if the adjustment increases the prediction error or slows model convergence. This reward mechanism guides the agent to learn in the direction of optimizing prediction performance, continuously improving its action strategy.

[0082] The agent's policy network and value network are updated using a deep deterministic policy gradient algorithm to optimize action selection. The policy network generates specific actions, while the value network evaluates the expected benefits of those actions. Both networks continuously adjust their parameters using gradient descent. In the production process of a certain electronic component, when the reliability level of the data link decreases and data stability becomes abnormal, the policy network learns to increase the input weight of vibration data and increase the iteration frequency. The value network then verifies whether this action can improve prediction accuracy. Through repeated iterations, the agent gradually masters adjustment strategies adapted to different scenarios.

[0083] Based on the optimal action output by the agent, the input weights and iteration frequency of the prediction model are dynamically adjusted. For example, in the prediction of blast furnace temperature in a steel plant, when the data link reliability level is improved and the temperature data is stable, the optimal action output by the agent may be to reduce the weight of fuel flow data and decrease the iteration frequency, so that the model can reduce the computational load while ensuring the prediction effect, and achieve real-time optimization of prediction accuracy.

[0084] Example 5: In S5, when dynamically adjusting the input weights and iteration frequency of the prediction model using a reinforcement learning algorithm, multi-objective optimization of the prediction model is also included. Multi-objective optimization aims to balance multiple key indicators in the prediction process, making the model more aligned with actual needs in complex industrial scenarios.

[0085] Prediction accuracy, model running efficiency, and resource utilization are used as input parameters for the multi-objective optimization function. Prediction accuracy reflects the degree of agreement between the model's output and actual industrial data, directly affecting the usability of the prediction results. Model running efficiency reflects the time required for the model to complete a prediction, affecting the real-time nature of the prediction, and is particularly important in fast-paced industrial scenarios. Resource utilization involves the consumption of computing and storage resources during model operation, and is related to cost control in industrial production. These three parameters measure model performance from different dimensions and together constitute the core elements of multi-objective optimization.

[0086] The weighting coefficients for each parameter are set, and their values ​​are determined based on the real-time requirements of the industrial scenario. In scenarios where the accuracy of prediction results is extremely important, such as quality prediction in precision instrument manufacturing, the weighting coefficient for prediction accuracy will be set higher. In scenarios where real-time performance is more critical, such as fault warnings in real-time production line monitoring, the weighting coefficient for model operating efficiency will be increased accordingly. In small and medium-sized factories with limited computing resources, the weighting coefficient for resource utilization may be given greater weight. By adjusting the weighting coefficients, the multi-objective optimization function can adapt to the priority requirements of different scenarios.

[0087] A non-dominated sorting genetic algorithm is used to solve a multi-objective optimization function. This algorithm simulates the natural selection and genetic mechanisms in biological evolution, transforming the optimization problem of multiple parameters into a population evolution problem. In the initial stage, a set of model parameter combinations is randomly generated as the initial population, with each parameter combination corresponding to a possible solution. For each solution in the population, its performance in three dimensions—prediction accuracy, runtime efficiency, and resource utilization—is calculated according to the multi-objective optimization function, and the solutions are then ranked using a non-dominated sorting method. The core of non-dominated sorting is to determine the superiority relationship between solutions; if a solution performs no worse than another solution in all parameters, and at least one parameter is better, then that solution dominates the other solution. Through multiple rounds of sorting, the solutions in the population are divided into different levels, with higher-level solutions closer to the optimal solution.

[0088] Based on the sorting, crowding degree calculations are used to select superior solutions for crossover and mutation operations. Crossover combines the parameters of two parent solutions to generate new child solutions, preserving the superior characteristics of the parents. Mutation randomly changes the parameters of some solutions, increasing population diversity and preventing the algorithm from getting trapped in local optima. After multiple generations of evolution, solutions in the population gradually cluster towards the optimal region, ultimately yielding the Pareto optimal solution set. Each solution in the Pareto optimal solution set has unique advantages, achieving a performance balance under different parameter combinations. For example, some solutions have high prediction accuracy but high resource consumption, some run quickly but have slightly lower accuracy, and some consume fewer resources but have balanced overall performance.

[0089] The target solution is selected from the Pareto optimal solution set based on the real-time needs of the industrial scenario. The selection process needs to comprehensively consider the characteristics of the current production task, the availability of resources, and the specific requirements for the prediction results. For example, in a production line experiencing a sudden equipment failure, it is necessary to quickly determine the cause of the failure. In this case, a target solution with high operating efficiency would be selected from the Pareto optimal solution set. However, when formulating long-term production plans, a target solution with high prediction accuracy may be selected.

[0090] The model parameters corresponding to the target solution are used as the initial policy of the reinforcement learning agent. This initial policy provides the agent with a high-performance starting point, reducing its learning cost. Based on this, the policy is optimized using a deep deterministic policy gradient algorithm. This algorithm combines the perceptual capabilities of deep neural networks with the decision-making power of deterministic policy gradients, enabling the agent to continuously adjust its policy through interaction with the environment. The agent outputs corresponding actions based on the current state input, receives feedback through a reward function, and then updates the parameters of the policy and value networks using gradient descent. Through continuous optimization, the agent's policy gradually improves, enabling it to more accurately output the input weight adjustments and iteration frequency adjustments of the prediction model, thus maintaining good performance in dynamically changing industrial environments.

[0091] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0092] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An industrial data prediction method based on AI machine learning, characterized in that: Includes the following steps; S1: During the parameter configuration of the initial prediction model, dynamic data in the industrial production environment is acquired in real time through a multi-dimensional data acquisition unit. After the dynamic data is preprocessed in layers, the stability and integrity of the data input are judged. S2: When data input is unstable or incomplete, perform a joint analysis of the fluctuation range of data acquisition frequency and the cumulative duration of transmission delay to assess the reliability level of the data acquisition link; S3: Based on the evaluation results, the reliability level of the data acquisition link is divided into high reliability, medium reliability and low reliability, and the corresponding link optimization strategy is executed. S4: When the reliability level of the data acquisition link is medium reliability, perform time series feature extraction on the preprocessed industrial data to evaluate the stationarity and periodicity of the data distribution. S5: Based on the reliability level of the data acquisition link and the stability and periodicity of the data distribution, the input weights and iteration frequency of the prediction model are dynamically adjusted using reinforcement learning algorithms.

2. The industrial data prediction method based on AI machine learning according to claim 1, characterized in that: In S1, a multi-dimensional data acquisition unit is deployed in the industrial production environment, including vibration sensors, current sensors, and humidity sensors. Real-time data is collected from the sensors at preset intervals and recorded as a raw dataset. The raw dataset includes the acquisition time, vibration amplitude, current intensity, and ambient humidity. The data integrity is labeled, and the labeled data is divided into a training sample set and a validation sample set for building and validating an adaptive filtering model. An adaptive filtering algorithm is used to initialize the filtering coefficients. The adaptive filtering model is trained using the training sample set, and the filtering effect of the model is verified using the validation sample set. The multi-dimensional data acquisition unit collects vibration amplitude, current intensity, and ambient humidity data in the industrial production environment in real time to form dynamic input data. The dynamic input data is input into the trained adaptive filtering model, and the model outputs the data noise coefficient based on the current data. The noise coefficient of the data output by the adaptive filtering model is compared with the preset noise threshold. If the noise coefficient is greater than or equal to the preset noise threshold, the data input is unstable or incomplete, and a data cleaning instruction is triggered. If the noise coefficient is less than the preset noise threshold, the data input is stable and complete, and the subsequent process continues.

3. The industrial data prediction method based on AI machine learning according to claim 1, characterized in that: In S2, a frequency anomaly index is generated based on the fluctuation range of the data acquisition frequency. The method for obtaining the frequency anomaly index is as follows: The clock module deployed in the data acquisition unit records the time interval of each data acquisition in real time and calculates the frequency deviation value for each acquisition cycle. Specifically, the calculation method is: the difference between the standard time interval and the actual time interval for two consecutive acquisition cycles; the smoothed frequency deviation value is calculated using a sliding window averaging method, where the smoothing process involves superimposing the current window's frequency deviation value with the smoothing result of the previous window according to a preset weight; the obtained smoothed frequency deviation value is compared with a pre-set standard frequency deviation range to calculate the frequency anomaly index, where the standard frequency deviation range is the interval between the mean of historical frequency deviations under normal operating conditions and three times the standard deviation; when the smoothed frequency deviation value exceeds the upper limit of the standard frequency deviation range, the frequency anomaly index is calculated as: the smoothed frequency deviation value minus the upper limit of the standard frequency deviation range; when the smoothed frequency deviation value is lower than the lower limit of the standard frequency deviation range, the frequency anomaly index is calculated as: the lower limit of the standard frequency deviation range minus the smoothed frequency deviation value; where the frequency anomaly index is a quantitative indicator measuring the degree of fluctuation in the data acquisition frequency.

4. The industrial data prediction method based on AI machine learning according to claim 3, characterized in that: The delay impact index is generated based on the cumulative duration of transmission delay within a fixed period. The method for obtaining the delay impact index is as follows: The data transmission delay time for each data transmission within a specified time period is obtained and recorded as a time series dataset. The empirical mode decomposition method is used to perform intrinsic mode function decomposition on the time series dataset to obtain multiple intrinsic mode components of different frequencies. Calculate the energy percentage of each intrinsic mode component. The energy percentage is calculated as the ratio of the variance of a single intrinsic mode component to the sum of the variances of all components. By combining the energy proportion of each component with the corresponding delay time, the delay impact index is calculated. Specifically, it is the sum of the products of the energy proportion of each component and the corresponding delay time.

5. The industrial data prediction method based on AI machine learning according to claim 4, characterized in that: The frequency anomaly index and delay impact index are converted into a second feature vector. This second feature vector is used as the input to a deep belief network model. The deep belief network model uses the prediction of the reliability level label of the data acquisition link for each set of second feature vectors as the training objective and minimizes the classification error of all reliability level labels as the optimization direction. The deep belief network model is trained until the classification error converges to a preset range and then training stops. The reliability level of the data acquisition link is determined based on the model output.

6. The industrial data prediction method based on AI machine learning according to claim 5, characterized in that: In S3, based on the evaluation results, the reliability level of the data acquisition link is divided into high reliability, medium reliability, and low reliability, as follows: The reliability level assessment value of the acquired data acquisition link is compared with a preset dual threshold, which includes a first reliability threshold and a second reliability threshold, and the first reliability threshold is less than the second reliability threshold. The reliability level assessment value is then compared with the first reliability threshold and the second reliability threshold respectively. If the reliability level assessment value is greater than the second reliability threshold, it is classified as high reliability and the current data acquisition and transmission parameters are maintained; if the reliability level assessment value is greater than or equal to the first reliability threshold and less than or equal to the second reliability threshold, it is classified as medium reliability and the data acquisition frequency and transmission bandwidth are adjusted to ensure a balance between data real-time performance and integrity; if the reliability level assessment value is less than the first reliability threshold, it is classified as low reliability, the backup data link is immediately activated, and a link failure warning is issued.

7. The industrial data prediction method based on AI machine learning according to claim 6, characterized in that: In S4, a data distribution characteristic index is generated based on the preprocessed industrial data to assess the stationarity and periodicity of the data distribution. The method for obtaining the data distribution characteristic index is as follows: Preprocessed industrial data within a fixed time period is acquired, and the data is detrended to eliminate the influence of long-term trends. The window size and smoothing coefficient of the local weighted regression model are determined, and the local weighted regression model is used to fit the detrended data to obtain the local trend curve of the data. The deviation between the actual data value and the local trend curve is calculated, and the absolute values ​​of the deviation values ​​are accumulated as the data distribution characteristic index. The data distribution characteristic index is a comprehensive indicator that measures the degree of deviation of the data from the local trend. The smaller the value, the more stable the data distribution.

8. The industrial data prediction method based on AI machine learning according to claim 7, characterized in that: The obtained data distribution characteristic index is compared with a pre-set data stability threshold. If the data distribution characteristic index is greater than or equal to the pre-set data stability threshold, the data distribution stability is low, and a data stability abnormal signal is generated. If the data distribution characteristic index is less than the pre-set data stability threshold, the data distribution stability is high, and a data stability normal signal is generated.

9. The industrial data prediction method based on AI machine learning according to claim 1, characterized in that: In S5, reinforcement learning algorithms are used to dynamically adjust the input weights and iteration frequency of the prediction model based on the reliability level of the data acquisition link and the stability and periodicity of the data distribution. The reliability level assessment value of the data acquisition link and the data distribution characteristic index are used as the state input of the reinforcement learning agent, and the input weight adjustment amount and iteration frequency adjustment amount of the prediction model are used as the action output of the reinforcement learning agent. The reward value of the current action is calculated through a preset reward function, with the reduction in prediction error and the model convergence speed as the core indicators. The agent's policy network and value network are updated using a deep deterministic policy gradient algorithm to optimize the action selection policy. Based on the optimal action output by the agent, the input weights and iteration frequency of the prediction model are dynamically adjusted to achieve real-time optimization of prediction accuracy.

10. The industrial data prediction method based on AI machine learning according to claim 9, characterized in that: In S5, when dynamically adjusting the input weights and iteration frequency of the prediction model using reinforcement learning algorithms, multi-objective optimization of the prediction model is also included. Specifically, prediction accuracy, model running efficiency, and resource utilization are used as input parameters of the multi-objective optimization function. Weight coefficients of each parameter are set, and the multi-objective optimization function is solved using a non-dominated sorting genetic algorithm to obtain the Pareto optimal solution set. Based on the real-time requirements of the industrial scenario, a target solution is selected from the Pareto optimal solution set, and the model parameters corresponding to the target solution are used as the initial policy of the reinforcement learning agent. Then, the policy is optimized using a deep deterministic policy gradient algorithm.

Images