Risk assessment method based on statistical model optimization

Abstract

The invention relates to the technical field of risk assessment optimization, and discloses a risk assessment method based on statistical model optimization. The method comprises the following steps: acquiring a running state data set of a target object containing multi-dimensional monitoring index time sequence data; and inputting the operation state data set into a pre-trained reference risk assessment model, and generating an initial risk score and risk distribution characteristics. And iteratively adjusting parameters of the reference risk assessment model through a dynamic correction algorithm according to the generated risk distribution characteristics, and generating an optimized risk assessment model. And adopting the optimized risk assessment model to re-assess the same operation state data set, and outputting a corrected risk score and a key risk area identifier. According to the method, the risk distribution characteristics of the specific data of the target object under the reference model are analyzed, and the model parameters are dynamically adjusted, so that the risk assessment standard better fits the actual risk mode, the accuracy and pertinence of risk identification are improved, and a more reliable basis is provided for a risk management and control decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of risk assessment optimization, and particularly to a risk assessment method based on statistical model optimization. BACKGROUND

[0002] Risk assessment is a key technology for ensuring the safe and stable operation of complex systems and is widely used in finance, industry, medicine, and other fields. Traditional risk assessment methods rely on rule systems constructed by expert experience or statistical models based on historical data. The rule system method assesses risk by setting a series of "if-then" logical judgments, and its effectiveness is highly dependent on the completeness and accuracy of expert knowledge, making it difficult to adapt to complex and changing environments. The statistical model method uses historical data to train a model to predict the probability of future risk occurrence, but the model performance is limited by the quality and representativeness of the training data.

[0003] Pre-trained risk assessment models face adaptability challenges in actual deployment. These models are usually trained on specific historical data sets, and their internal assumptions and parameter settings are closely related to the distribution of the training data. When applied to new target objects or when the operating environment changes, the evaluation effect of the model will often decrease due to the shift in data distribution, resulting in scoring bias or misjudgment. For example, in industrial equipment monitoring, the normal and abnormal state indicator thresholds of the same type of equipment may shift at different operating conditions and different wear stages, and a model with fixed parameters cannot accurately capture this change.

[0004] Existing risk assessment systems lack effective model self-optimization mechanisms. When the evaluation results deviate significantly from the actual risk situation, manual intervention is usually required to reselect features, adjust thresholds, or even retrain the model. This process is time-consuming and labor-intensive, and relies on professional data analysts, making it difficult to respond to dynamically changing risk situations in a timely manner. In addition, the risk score output by traditional methods is usually a single numerical value or grade, lacking in-depth analysis of the risk space distribution characteristics, and unable to clearly indicate the main sources of risk from which monitoring indicators or system links, limiting the targetedness of risk control measures.

[0005] There is an urgent need for a new risk assessment method that can automatically adapt to the characteristics of target objects, dynamically optimize model parameters, and provide detailed risk distribution characteristics. SUMMARY

[0006] The present application aims to provide a risk assessment method based on statistical model optimization to solve the problems raised in the background.

[0007] To achieve the above-mentioned purpose, the present application provides a risk assessment method based on statistical model optimization, which comprises: acquire a running state data set of a target object, the running state data set containing time series data of multidimensional monitoring indicators; input the running state data set into a pre-trained benchmark risk assessment model to generate an initial risk score and a risk distribution feature; According to the risk distribution feature, the parameters of the benchmark risk assessment model are iteratively adjusted by a dynamic correction algorithm to generate an optimized risk assessment model; The optimized risk assessment model is used to re-evaluate the running state data set, and a corrected risk score and a key risk area identifier are output.

[0008] Preferably, the running state data set of the target object is acquired, comprising: Collect the running parameters of the target object within a preset time range, the running parameters including mechanical vibration data, temperature change data and load fluctuation data; The mechanical vibration data is subjected to frequency spectrum decomposition processing to extract characteristic frequency components and corresponding amplitude variation curves; The temperature change data and the load fluctuation data are aligned according to the time stamp to generate a temperature-load coupling sequence.

[0009] Preferably, the running state data set is input into a pre-trained benchmark risk assessment model, comprising: The characteristic frequency components are input into the first analysis layer of the benchmark risk assessment model to calculate the vibration risk coefficient by frequency domain feature matching algorithm; The temperature-load coupling sequence is input into the second analysis layer of the benchmark risk assessment model to generate the thermal risk coefficient by nonlinear regression analysis; The vibration risk coefficient and the thermal risk coefficient are fused to generate the initial risk score and the corresponding risk distribution feature.

[0010] Preferably, the parameters of the benchmark risk assessment model are iteratively adjusted by a dynamic correction algorithm, comprising: Extract the abnormal fluctuation interval in the risk distribution feature, calculate the duration and fluctuation amplitude of the abnormal fluctuation interval; According to the ratio of the duration and fluctuation amplitude, the correction weight of the benchmark risk assessment model is determined; Based on the correction weight, the output layer parameters of the benchmark risk assessment model are updated by gradient.

[0011] Preferably, the running state data set is re-evaluated by the optimized risk assessment model, comprising: inputting the characteristic frequency component into a first analysis layer of the optimized risk assessment model, and calculating an updated vibration risk coefficient by using a corrected frequency domain feature matching algorithm; inputting the temperature-load coupling sequence into a second analysis layer of the optimized risk assessment model, and generating an updated thermal risk coefficient by using an adjusted nonlinear regression analysis; fusing the updated vibration risk coefficient and the updated thermal risk coefficient, and outputting a corrected risk score and a key risk area identifier.

[0012] Preferably, the generation process of the key risk area identifier comprises: extracting a high-frequency abnormal component in the updated vibration risk coefficient, and marking a mechanical vibration high-risk area; extracting a temperature mutation interval in the updated thermal risk coefficient, and marking a thermal impact high-risk area; superimposing and analyzing spatial distribution data of the mechanical vibration high-risk area and the thermal impact high-risk area, and generating the key risk area identifier.

[0013] Preferably, the method further comprises: obtaining a historical maintenance record of the target object, and extracting a fault occurrence time and a fault type in the historical maintenance record; matching the fault occurrence time with a time stamp of the key risk area identifier, and calculating a fault prediction accuracy rate; secondarily correcting parameters of the optimized risk assessment model according to the fault prediction accuracy rate.

[0014] Preferably, the secondary correction of the parameters of the optimized risk assessment model comprises: if the fault prediction accuracy rate is lower than a preset threshold, increasing a calculation weight of the vibration risk coefficient; if the fault prediction accuracy rate is higher than the preset threshold, decreasing a calculation weight of the thermal risk coefficient; regenerating the corrected risk score and the key risk area identifier based on the adjusted weight coefficient.

[0015] Preferably, the method further comprises: collecting real-time running data of the target object, inputting the real-time running data into the secondarily corrected risk assessment model, and generating a real-time risk score; when the real-time risk score exceeds a dynamic warning threshold, triggering a risk warning signal; generating a corresponding risk mitigation strategy according to the risk warning signal.

[0016] Preferably, the determination process of the dynamic warning threshold comprises: statistical history running data, calculate the mean and standard deviation of the risk score; According to the linear combination of the mean and standard deviation, the alarm threshold is dynamically adjusted; When the fluctuation amplitude of the real-time running data increases, the dynamic alarm threshold is reduced.

[0017] Compared with the prior art, the beneficial effects of the present application are: The present application effectively improves the adaptability of the model to new scenarios by dynamically correcting the pre-trained model using the running state data of the target object. The baseline model may be trained based on general data, and its evaluation standard may not completely match the actual risk characteristics of the specific target object. This method intelligently adjusts the model parameters by analyzing the risk distribution characteristics of the target object data under the baseline model, making the evaluation criteria more consistent with the real risk pattern of the object, and reducing the evaluation deviation caused by data distribution differences.

[0018] The dynamic correction algorithm realizes the automatic and iterative optimization of model parameters, reducing the dependence on manual intervention. Traditional model optimization usually requires offline retraining, which is cumbersome. This method can automatically and continuously fine-tune model parameters based on the characteristic feedback of real-time evaluation results during online model operation, enabling the model to have self-learning and adaptive capabilities, keeping up with changes in the target object state and maintaining the timeliness of evaluation.

[0019] The introduction of risk distribution characteristics expands risk assessment from a single scalar output to multi-dimensional feature analysis. This method not only gives the final risk score, but also analyzes the distribution of risk in multiple monitoring dimensions. This feature analysis helps to understand the composition of risk and identify key indicators or system weak links that contribute most to the overall risk, providing more comprehensive information for precise risk management.

[0020] The corrected risk score and key risk area identification output by the optimized model provide more reliable and specific guidance for risk response decisions. The corrected score more accurately reflects the current actual risk level, while the key area identification directly indicates the focus direction that needs priority attention and intervention measures. This enables risk management resources to be more effectively allocated, improving the efficiency and effectiveness of risk control.

[0021] This method establishes a closed-loop process from evaluation to feedback to optimization, realizing the continuous self-improvement of the risk assessment system. The system can gradually approach the optimal evaluation state by continuously comparing the differences between the evaluation results before and after optimization and adjusting the model accordingly, forming a virtuous cycle of getting better and better, and improving the long-term performance and practical value of the risk assessment system. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 for optimizing the front and back risk score comparison chart; Figure 2 for obtaining the running state data set of the target object; Figure 3 for inputting the running state data set into the pre-trained benchmark risk assessment model; Figure 4 for risk assessment coefficient and score optimization comparison chart. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] Please refer to Figure 1 The present application provides a risk assessment method based on statistical model optimization, which comprises: obtaining the running state data set of the target object, inputting the running state data set into the pre-trained benchmark risk assessment model to generate the initial risk score and risk distribution characteristics, iteratively adjusting the parameters of the benchmark risk assessment model according to the risk distribution characteristics by using a dynamic correction algorithm, thereby generating an optimized risk assessment model, finally re-evaluating the running state data set by using the optimized risk assessment model, and outputting the corrected risk score and the key risk area identification. The running state data set contains time series data of multi-dimensional monitoring indicators, which can fully reflect the running health status of the target object. The pre-trained benchmark risk assessment model is trained based on historical normal running data and has preliminary risk identification ability. The dynamic correction algorithm focuses on the model bias exposed in the risk distribution characteristics, and improves the adaptability and evaluation accuracy of the model to the current running state through parameter iterative adjustment. The optimized model not only outputs more accurate corrected risk score, but also locates the key risk area, providing clear guidance for targeted maintenance.

[0025] Embodiment 1: refer to Figure 2, the construction of the operating state data set begins with the comprehensive collection of the operating parameters of the target object within a preset time range. The setting of the preset time range needs to cover the complete working cycle of the target object and various typical working conditions that may occur, such as starting, steady-state operation, load change, and shutdown stage. The operating parameters include mechanical vibration data, temperature change data, and load fluctuation data, which are synchronously obtained from different sensors and monitoring units. The mechanical vibration data is collected by piezoelectric acceleration sensors arranged on key mechanical components such as bearings, gearboxes, or housings. The sampling frequency of the sensors is set to be more than twice the highest vibration frequency that may occur to satisfy the Nyquist sampling theorem, thereby ensuring the integrity of the vibration signal information. The temperature change data is measured by embedded thermocouples or non-contact infrared thermometers, and the selection of temperature measurement points covers the heat source area and key nodes on the heat conduction path. The load fluctuation data is directly read from the power output unit of the motor driver, the pressure sensor of the hydraulic system, or the load command signal of the control system. The accuracy of the load data directly affects the reliability of subsequent analysis. Spectral decomposition processing of the mechanical vibration data is a key step in extracting feature information. The spectral decomposition processing converts the time-domain vibration signal into a frequency-domain representation using the Fast Fourier Transform algorithm. The Fast Fourier Transform processing decomposes the time series vibration waveform into a series of discrete frequency components, each corresponding to a vibration mode or excitation source. The extraction of characteristic frequency components is based on known theoretical characteristic frequencies of the rotating components of the target object, such as rotational speed, gear meshing frequency, and bearing passing frequency. By setting a frequency window around these theoretical frequencies, amplitude integration or peak searching is performed. The corresponding amplitude variation curve is generated by tracking the amplitude evolution of each characteristic frequency component over time within the preset time range. The amplitude variation curve can reflect the development trend of component wear, imbalance aggravation, or looseness. The spectral decomposition processing not only includes the fundamental frequency component but also focuses on the harmonic component and the sideband structure. These frequency characteristics often contain rich fault diagnosis information. Aligning the temperature change data and the load fluctuation data by timestamp is a prerequisite for constructing the temperature-load coupling sequence. The timestamp alignment operation uses a unified clock source to accurately mark the time of all monitoring data. The data alignment process needs to solve the problems of different sensor data collection start time differences and non-synchronous sampling periods. Usually, an interpolation algorithm is used to resample the non-uniformly sampled data sequence to a unified time axis. The essence of generating the temperature-load coupling sequence is to establish the corresponding relationship sequence between the temperature reading and the load value at each time point. This sequence exists in the form of two-dimensional data points (load value, temperature value) arranged in chronological order. The temperature-load coupling sequence can intuitively show the hysteresis effect of load change on temperature, the nonlinear temperature rise process, and the thermal equilibrium state, such as the instantaneous response speed and steady-state temperature difference of temperature under high load sudden working conditions.

[0026] The quality of mechanical vibration data directly affects the effect of frequency spectrum analysis. Before frequency spectrum decomposition, the original vibration signal usually needs to be preprocessed. The preprocessing operation includes removing the DC offset, applying the Hanning window or Hamming window function to reduce the frequency spectrum leakage, and performing band-pass filtering to eliminate the interference of high-frequency noise and low-frequency drift. The identification of characteristic frequency components requires the combination of device structure parameters and operating conditions. For example, for devices operating at variable frequency, the characteristic frequency will drift with the change of rotating speed. At this time, the order ratio analysis technique is needed to convert the frequency spectrum diagram into an order ratio spectrum relative to the rotating speed to stabilize the position of the characteristic frequency. The coupling analysis of temperature change data and load fluctuation data needs to consider the time delay effect caused by thermal inertia. Simple timestamp alignment may not fully reflect the dynamic relationship between load and temperature. Therefore, when generating temperature-load coupling sequences, a time offset compensation algorithm is sometimes introduced, that is, according to the heat transfer model, the delay time of temperature response to load change is estimated, and the load data is advanced by the corresponding delay amount on the time axis before being aligned with the temperature data. This dynamic alignment method can more accurately capture the direct cause of load change leading to temperature change, so that the temperature-load coupling sequence can better reflect the thermal dynamic characteristics of the system. The storage format of the sequence uses a structured array or a time series database, and each data point contains a timestamp, a load value, a temperature value, and possibly a data quality identifier.

[0027] The final integration of the operating state data set will collect the mechanical vibration features processed as described above and the temperature-load coupling sequence. The data set is organized in the form of multi-dimensional time series, with each dimension representing a monitoring indicator. All dimensions share a unified time coordinate axis. The metadata part of the data set records the parameters of the acquisition device, sensor number, sampling frequency, preprocessing method, etc. Information ensures data traceability and repeatability of the analysis process. The complete operating state data set constitutes the data basis for the initial assessment of the subsequent benchmark risk assessment model, and its quality and integrity are directly related to the accuracy of the risk assessment result. The data acquisition and processing process is usually realized by embedded data acquisition cards and upper computer software, realizing automatic data acquisition, real-time processing and local storage.

[0028] Embodiment 2: see Figure 3The characteristic frequency component is input to the first analysis layer of the benchmark risk assessment model, and the core of the first analysis layer is a frequency domain feature matching algorithm. The frequency domain feature matching algorithm operates in dependence on a pre-generated fault characteristic frequency library, which stores the characteristic frequency calculation formulas and historical normal ranges corresponding to various typical fault modes. The algorithm execution process is to compare the real-time input characteristic frequency component with each entry in the fault characteristic frequency library one by one. The comparison operation calculates the deviation of the amplitude of the real-time frequency component from the historical normal amplitude benchmark of the corresponding frequency in the library. This deviation is quantified by calculating the Euclidean distance or Mahalanobis distance in a mathematical manner. The calculation result is normalized to a value between zero and one, which is defined as the vibration risk coefficient. The closer the vibration risk coefficient is to one, the higher the matching degree between the real-time vibration feature and a certain fault feature, and the higher the risk level of the mechanical state. The temperature-load coupling sequence is input to the second analysis layer of the benchmark risk assessment model, and the processing core of the second analysis layer is nonlinear regression analysis. Nonlinear regression analysis aims to establish a mathematical model that can describe the change of temperature with load under normal working conditions. This model can be a polynomial model, an exponential model, or a regressor based on machine learning. The model is trained using historical normal data, with load fluctuation data as input and temperature prediction value as output. In actual application, the nonlinear regression analysis inputs the load data in the real-time collected temperature-load coupling sequence into the trained regression model to calculate the corresponding temperature prediction value sequence. The temperature prediction value sequence is compared with the real-time collected real temperature value sequence to calculate the residual error, i.e., the difference between the predicted value and the real value. The statistical characteristics of the residual error sequence (such as the root mean square value of the residual error and the maximum absolute residual error) are used to calculate the thermal risk coefficient. Larger residual error indicates that the temperature rise behavior of the target deviates from the normal pattern learned by the model, which may indicate poor heat dissipation, increased friction, or internal heat source abnormalities, and thus the thermal risk coefficient is higher.

[0029] The initial risk score and risk distribution features are generated by fusing the vibration risk coefficient and the thermal risk coefficient. The fusion strategy adopts weighted linear superposition or nonlinear mapping based on fuzzy logic. The weighted linear superposition assigns fixed weights to the vibration risk coefficient and the thermal risk coefficient, and the weight ratio is determined based on domain knowledge. For example, the weight of the vibration risk of the rotating machinery is usually higher than that of the thermal risk. The result of the weighted summation is converted by a Sigmoid function to output an initial risk score between zero and one hundred. The generation of the risk distribution features requires more detailed data. The initial risk score is a global evaluation at a moment, while the risk distribution features need to show the distribution of the risk on the time axis and the frequency axis. The risk distribution features can be a two-dimensional matrix, where the rows correspond to different characteristic frequency bands, the columns correspond to different time segments, and the matrix elements represent the risk contribution of the frequency band in the time segment. Another risk distribution feature can be a one-dimensional sequence about time, where each point in the sequence contains the vibration risk coefficient, the thermal risk coefficient, and the local risk value after fusion, which reflects the trend of the risk evolution over time.

[0030] The parameters of the benchmark risk assessment model are iteratively adjusted by a dynamic correction algorithm, and the trigger condition of the dynamic correction algorithm is the generation of risk distribution features. The first step of the dynamic correction algorithm is to extract the abnormal fluctuation interval in the risk distribution features, which refers to the data interval in which the risk values at multiple consecutive time points are significantly higher than the overall average level in the time series of the risk distribution features. Identifying the abnormal fluctuation interval requires setting a dynamic threshold, for example, using a moving average line plus twice the moving standard deviation as the threshold line. The continuous data segment exceeding the threshold line is marked as an abnormal fluctuation interval. For each identified abnormal fluctuation interval, two key indicators need to be calculated: duration and fluctuation amplitude. The duration is the length of the abnormal fluctuation interval on the time axis, and the fluctuation amplitude is the integral or maximum value of the average excess of the threshold value in the interval. The correction weight of the benchmark risk assessment model is determined according to the ratio of the duration and the fluctuation amplitude, which is called the abnormal intensity index. The greater the abnormal intensity index, the more persistent and intense the abnormal state, and therefore the greater the correction of the benchmark risk assessment model. The correction weight is a coefficient between zero and one, which is positively correlated with the abnormal intensity index and can be obtained by a pre-set linear or non-linear function mapping. Based on the correction weight, the output layer parameters of the benchmark risk assessment model are updated by gradient, where the output layer parameters specifically refer to the weights and bias parameters responsible for mapping intermediate features to the final risk score. The gradient update adopts the principle of gradient descent, but the loss function is defined as the difference between the risk score output of the current model for the abnormal fluctuation interval and a higher target risk value. The correction weight is used to scale the gradient value calculated by the loss function, thereby controlling the step size of parameter update. The greater the correction weight, the greater the step size of parameter update, and the greater the adjustment of the model, making it more sensitive to similar abnormal fluctuation patterns in the future. The entire dynamic correction process is an iterative loop, and each new running state data set input and new risk distribution feature generation may trigger a new round of parameter adjustment, so that the benchmark risk assessment model can continuously adapt to the changes in the state of the target object.

[0031] The pre-trained benchmark risk assessment model itself is a statistical model with a multi-layer structure, and the training data of the model comes from the data records of the target object in a healthy state during long-term historical operation. The training process enables the benchmark risk assessment model to learn the multi-parameter correlation pattern in the normal state. The frequency domain feature matching algorithm of the first analysis layer includes not only the theoretically calculated frequency, but also the actual feature frequency and its evolution law extracted from the analysis of historical minor fault cases. The nonlinear regression analysis model of the second analysis layer needs to select the appropriate function form and hyperparameters in the training stage to best fit the nonlinear relationship between load and temperature, avoiding overfitting or underfitting. The setting of the weight in the model fusion stage is not fixed, and the initial weight can be determined based on the fault mode influence analysis and adjusted as a parameter in the dynamic correction process. The extraction accuracy of the risk distribution characteristics depends on the resolution of the data. High sampling rate data helps to capture transient abnormal pulses, but also puts higher requirements on computing and storage resources. The gradient update operation in the dynamic correction algorithm needs to consider the setting of the learning rate. A too large learning rate may cause model oscillation, and a too small learning rate may cause slow correction speed. In the actual embedded system, the dynamic correction algorithm may run in a timed batch processing mode rather than updating immediately for each data point, to balance the computing load and real-time requirements. The version management of the benchmark risk assessment model is also an aspect that needs to be considered. Important model parameter versions need to be archived to allow rollback to the previous stable state when the correction effect is not ideal.

[0032] Referring to Figure 4 In the risk assessment method based on statistical model optimization, the figure integrates the dynamic evolution process of risk coefficient calculation and score optimization. In specific operation, the risk coefficient calculation result graph presents the normalized time series of the vibration risk coefficient (black curve) and the thermal risk coefficient (gray curve), where the vibration risk coefficient is calculated by the frequency domain feature matching algorithm, comparing the real-time feature frequency component with the historical benchmark deviation of the fault feature frequency library, and quantifying and normalizing the Euclidean distance; the thermal risk coefficient is generated by nonlinear regression analysis, based on the residual statistical characteristics of the temperature-load coupling sequence. Both of them have peak values near time point 3, with the vibration risk coefficient reaching 0.8 and the thermal risk coefficient approaching 0.6, reflecting the abnormality of multi-dimensional risk superposition. The risk score optimization comparison chart shows the evolution of the initial risk score (dark gray solid line) and the optimized score (light gray solid line) relative to the warning threshold (dashed line). The initial score is generated by weighted fusion of risk coefficients, and the optimized score is adjusted by the dynamic correction algorithm to iteratively adjust the model parameters. The score is reduced by about 20% and the variance is reduced, indicating that the model correction effectively suppresses the risk. The entire chart correlates the risk identification and optimization process through a unified time axis, and intuitively reveals the role of the dynamic correction algorithm in balancing model sensitivity and stability.

[0033] The optimized risk assessment model has completed the iterative adjustment of parameters, and its internal structure has changed compared to the benchmark risk assessment model. The characteristic frequency component is input into the first analysis layer of the optimized risk assessment model, and the modified frequency domain feature matching algorithm executed by the first analysis layer is consistent in core logic with the benchmark model, but the key parameters have been updated. The fault characteristic frequency library relied on by the modified frequency domain feature matching algorithm may have added new fault mode features or adjusted the reference amplitude threshold of existing characteristic frequencies. When calculating the matching degree of real-time characteristic frequency components and library characteristic frequencies, the algorithm may use a different similarity measurement method, such as changing from Euclidean distance to Mahalanobis distance to consider the correlation between features. The matching degree calculation result will be multiplied by a sensitivity gain coefficient generated by a dynamic correction process before being normalized into a vibration risk coefficient, which directly amplifies the response to abnormalities in a specific frequency range. The updated vibration risk coefficient thus better reflects the key vibration risks identified through the model optimization stage.

[0034] The temperature-load coupling sequence is input into the second analysis layer of the optimized risk assessment model, and the adjusted nonlinear regression analysis run by the second analysis layer has modified the structure or coefficients of its regression model. The adjusted nonlinear regression analysis may extend the original second-order polynomial model to a third-order model to better capture the steeper change relationship between load and temperature in the high temperature range. The input features of the regression model may be expanded, with the load change rate in addition to the real-time load value as an input to describe the dynamic thermal response of the system. The loss function used during model training may be changed from mean square error to mean absolute error, making the model more robust to abnormal temperature points. The mapping function parameters used when converting the calculated residual into a thermal risk coefficient have also been adjusted, such as lowering the residual threshold for triggering a high risk score, making the model more sensitive to slight thermal abnormalities. The updated thermal risk coefficient thus contains more information about the thermal state. The updated vibration risk coefficient and the updated thermal risk coefficient are fused to output a modified risk score, and the fusion function F used in the fusion process is as follows: Where: represents the modified risk score, represents the updated vibration risk coefficient, represents the updated thermal risk coefficient. The function and are introduced monotonic nonlinear transformation functions for stretching or compressing the risk coefficient, thereby adjusting its contribution characteristics in the final score. Coefficients and are fusion weights, and satisfy The two weights are recalibrated according to the risk distribution characteristics in the model optimization stage. The modified risk score The value range of the modified risk score is usually set between zero and one hundred, and the higher the score represents the higher the comprehensive risk level. Compared with the initial risk score, the value of the modified risk score more accurately reflects the comprehensive judgment of the current operating state after model modification.

[0035] The generation of the key risk area identification starts with the in-depth analysis of the updated vibration risk coefficient, which is a sequence that changes over time. Extracting high-frequency abnormal components in the updated vibration risk coefficient requires the application of signal processing techniques. High-frequency abnormal components refer to the rapidly changing, high-frequency fluctuation components in the vibration risk coefficient sequence. A digital high-pass filter is used to filter the vibration risk coefficient sequence, filtering out the slowly changing trend items and retaining the rapidly fluctuating details. In the filtered sequence, the high-frequency fluctuation period with an amplitude exceeding the preset threshold is marked as a mechanical vibration high-risk area. These areas usually correspond to transient impact events, short-term resonance phenomena or intermittent jamming between components, etc. Precursors of transient failures. Extracting temperature mutation intervals in the updated thermal risk coefficient focuses on the dynamic characteristics of thermal risk. Temperature mutation intervals refer to sections in the thermal risk coefficient sequence that exhibit rapid stepwise increases or decreases. Identifying temperature mutation intervals involves calculating the first-order difference of the thermal risk coefficient using a sliding window. When the absolute value of the difference exceeds the set mutation threshold for consecutive points, the interval is determined to be a temperature mutation interval. Temperature mutation intervals are marked as thermal impact high-risk areas, and thermal impact may cause material thermal fatigue, seal failure or lubrication performance degradation. The marking operation not only records the time point of the mutation, but also records the amplitude and duration of the mutation, forming an area identification with risk intensity information.

[0036] The key to generating the comprehensive risk view is to overlay the spatial distribution data of the mechanical vibration high-risk area and the thermal impact high-risk area. The spatial distribution data contains the physical location information of each high-risk area, such as for large equipment, the location information can be accurate to a specific bearing seat number, gear box level or cooler unit. The overlay analysis is carried out in the time-space two-dimensional coordinate system, and for the same physical location, it is checked whether the mechanical vibration high-risk area and the thermal impact high-risk area exist in the time dimension. The overlay analysis adopts the logical "and" operation, that is, only when the same location is marked as both high-risk areas in the same time period, the location is determined as a key risk area. For the areas with time overlap but adjacent in space, the risk correlation strength is calculated according to the distance decay model, and if the strength exceeds the threshold, the areas are merged and identified. The final generated key risk area identification is a list or map, and each item in the list records the spatial location, risk type combination, risk strength level and time window information of the risk area. The key risk area identification provides precise spatial positioning guidance for maintenance personnel's on-site inspection and intervention measures.

[0037] The calculation process of the optimized risk assessment model is embedded in the data processing system, which obtains the preprocessed feature frequency components and temperature-load coupling sequences from the data bus in real time. The updated vibration risk coefficient and the updated thermal risk coefficient calculated by the model are temporarily stored in the circular buffer area for the fusion module to call. The fusion weight coefficient and is stored as a model parameter in the configuration file, which can be flexibly set according to different equipment types or operation and maintenance strategies. The specific form of the nonlinear transformation function and is usually a lookup table function or a piecewise linear function to emphasize the key risk coefficient interval. The generation algorithm of the key risk area identification periodically scans the risk coefficient sequence, and the filter type, cutoff frequency and window length need to be carefully selected to avoid introducing phase distortion or missing reports. The overlay analysis of the spatial distribution data needs to access the structured three-dimensional model database of the equipment to map the abstract data points to specific physical components. The whole generation process of the key risk area identification emphasizes the correlation analysis of multi-source risk information in the time-space dimension, thereby revealing the systematic risk hotspots that cannot be found by single-dimensional analysis. The generated identification information is presented to the equipment operation and maintenance personnel in the form of highlights, flashes or list alarms through the human-computer interface.

[0038] Example 4: Referring to Table 1, obtaining the historical maintenance records of the target object is the basis for model performance verification. The historical maintenance records are stored in the enterprise's computerized maintenance management system and contain structured event logs. Extracting the failure occurrence time and failure type from the historical maintenance records requires accessing specific fields in the database. The failure occurrence time is accurate to the day, hour, and minute, and the failure type follows a unified classification coding system. The failure type coding includes major categories such as mechanical failure, electrical failure, sensor failure, and thermal failure, with more detailed subcategories under each major category. The historical maintenance records may also contain text descriptions of failure phenomena, repair measures taken, replacement parts information, and downtime, which can help better understand the nature of the failure. The data extraction process requires data cleaning to address issues such as record missing, timestamp errors, or inconsistent type coding, ensuring the accuracy of subsequent analysis. Matching the failure occurrence time with the timestamp of the key risk area identification is the core step in calculating the failure prediction accuracy. The timestamp of the key risk area identification records the specific time when the risk area is identified. The matching operation aims to find whether the key risk area identification has occurred before the historical failure. The matching rule sets a warning time window, for example, 24 hours to 1 hour before the failure occurs. As long as the timestamp of the key risk area identification falls within this window, it is considered a successful prediction. The matching process needs to address the accuracy of time alignment. The timestamps of the device data collection system and the maintenance record system must be synchronized and corrected based on a unified clock source. The matching result generates a list, with each item in the list recording a historical failure event and the matching key risk area identification information. Calculating the failure prediction accuracy requires statistics based on the matching results. The failure prediction accuracy is defined as the ratio of the number of successfully predicted failures to the total number of failures. The number of successfully predicted failures refers to the number of failure events that have key risk area identification matching within the warning time window. The total number of failures is the sum of all failure events that occurred within the selected historical analysis period and have complete records available for analysis. The calculation formula of the failure prediction accuracy P is the number of successfully predicted failures divided by the total number of failures multiplied by 100, resulting in a percentage value. The calculation process needs to consider the independence of failure events. For multiple related failures caused by the same root cause, they may be considered as a composite failure event when counted.

[0039] Table 1: Failure prediction accuracy analysis table According to the fault prediction accuracy, the parameters of the optimized risk assessment model are secondarily corrected. The secondary correction strategy is based on the comparison result of the fault prediction accuracy and the preset threshold. The preset threshold is an empirical value, for example, seventy percent, which represents the minimum acceptable prediction performance standard. If the fault prediction accuracy is lower than the preset threshold, and the analysis of the missed fault event type shows that the model is not sensitive to mechanical vibration related faults, the calculation weight of the vibration risk coefficient is increased. The calculation weight of the vibration risk coefficient is increased by modifying the weight parameter in the model fusion formula, for example, increasing the weight a of the vibration risk coefficient by 0.1, while correspondingly reducing the weight β of the thermal risk coefficient, keeping the total weight sum to 1. The adjustment range of the weight can be proportional to the degree that the fault prediction accuracy is lower than the threshold. If the fault prediction accuracy is higher than the preset threshold, indicating that the overall performance of the model is good, but analyzing the successful prediction and false alarm cases may find that there is excessive warning for thermal risk, the calculation weight of the thermal risk coefficient is reduced. The calculation weight of the thermal risk coefficient is reduced by reducing the weight β of the thermal risk coefficient in the fusion formula, for example, reducing it by 0.05, and adding the reduced weight value to the weight a of the vibration risk coefficient. This adjustment aims to reduce non-critical thermal warnings, so that the model focuses more on the main mechanical vibration risk and improves the pertinence of the warning. The adjustment of the weight coefficient is recorded in the configuration file of the model as a new version parameter of the model.

[0040] Based on the adjusted weight coefficients, the corrected risk score and the key risk area identification are regenerated, and this process is to re-execute the risk assessment process. The optimized risk assessment model uses the secondarily corrected new weight coefficients to re-evaluate the same historical running state data set or new real-time data. The calculation process of the vibration risk coefficient and the thermal risk coefficient remains unchanged, but the new weights a and β are used for weighted synthesis in the fusion stage. The corrected risk score recalculated due to the change of weight distribution may be different from the previous score. The score originally dominated by thermal risk may more clearly reflect the vibration risk. The generation logic of the key risk area identification remains unchanged, but the area extraction and superposition analysis are based on the updated risk coefficients. The newly generated key risk area identification may change in spatial distribution and temporal distribution, and some pure mechanical vibration high-risk areas previously hidden by thermal risk are revealed. The secondary correction process introduces an important feedback mechanism, which uses historical facts to calibrate the model, so that the risk assessment model has the ability to learn from actual operation effect, and gradually improves its practicality in actual application. The whole process embodies the closed loop of model parameter iteration optimization based on historical performance data, which helps to improve the reliability and credibility of the risk assessment system.

[0041] The data quality of historical maintenance records directly affects the reliability of the verification results, and the completeness, accuracy and timeliness of the maintenance records need to be guaranteed. There may be errors in the recording of the time of failure occurrence, such as the time of failure occurrence, the time of failure discovery and the time of maintenance record entry, and when analyzing, it needs to be clear which time point is taken as the benchmark. The accuracy of the time stamp of the key risk area identification is also very important, and high-frequency data collection will produce more accurate time stamps, which is conducive to more accurate matching. The accuracy of failure prediction is a macroscopic indicator, and further analysis of the prediction accuracy of various types of failures can be carried out to make more targeted model corrections. For large and complex equipment, secondary correction can be considered for subsystems or components, and different weight parameters can be set for different subsystems to achieve more refined model adjustment. The secondary correction process can be automatically executed periodically, such as once every quarter or every half year, and the new historical maintenance record data accumulated during the period is used to make the risk assessment model continuously adapt to the changes in the equipment state. The management of model versions is very important, and after each parameter correction, the old version of the parameter configuration should be saved to quickly return to the previous stable state when the new correction effect is not ideal. Based on the adjusted weight coefficients, the results of the corrected risk score and the identification of key risk areas are regenerated, which can be compared with the latest operating data to observe the changes in the model behavior and verify the actual effect of the secondary correction.

[0042] Example 5: Collecting real-time operation data of the target object is the basis for the continuous operation of the system. For centrifugal compressors, real-time operation data includes vibration acceleration signals of bearing housings, inlet and outlet temperatures of each stage of coolers, current and power data of motor driving ends. Vibration acceleration signals are collected by ICP type acceleration sensors installed on compressor rotor support bearings, with a sampling frequency of 10 kHz to capture possible high frequency impact components. Temperature data is measured by platinum resistance temperature sensors, with measurement points arranged in the compressor cylinder, intermediate cooler and lubricating oil system. Load fluctuation data is obtained by monitoring the output power signal of the motor frequency converter, which reflects the real-time changes of the gas load handled by the compressor. All these data are collected and analog-to-digital converted by distributed I / O modules arranged in the field of the equipment, and transmitted to the real-time database of the central monitoring system through industrial Ethernet. The data stream is updated at fixed time intervals to ensure that the risk assessment model can obtain the latest equipment state information. Real-time operation data is input into the risk assessment model after secondary correction to generate real-time risk scores. Real-time operation data needs to go through the same preprocessing process as historical data before inputting into the model. Fast Fourier transform is performed on the vibration acceleration signal to extract characteristic frequency components related to rotor speed frequency and gear meshing frequency and their amplitudes. Temperature data and power data are aligned by timestamp to generate real-time temperature-load coupling sequences. The preprocessed feature data is sent to the risk assessment model which has been optimized based on historical maintenance records and corrected. The first analysis layer of the model uses the corrected frequency domain feature matching algorithm to process vibration features and calculate real-time vibration risk coefficients. The second analysis layer of the model uses the adjusted nonlinear regression analysis to process the temperature-load coupling sequence and calculate real-time thermal risk coefficients. Finally, the fusion module uses the weight coefficients determined after secondary correction to combine the two risk coefficients into a real-time risk score between zero and one hundred. This score serves as a comprehensive indicator, updated every second, dynamically reflecting the instantaneous health status of the centrifugal compressor.

[0043] When the real-time risk score exceeds the dynamic warning threshold, a risk warning signal is triggered. The calculation of the dynamic warning threshold relies on historical operation data. The distribution of risk scores in historical operation data is statistically analyzed, and a longer period of historical data when the equipment is in normal stable operation is selected to calculate the mean and standard deviation of all risk scores in this period. The dynamic warning threshold is usually set to , and the coefficient is determined according to the sensitivity requirements of the system to risk, for example =2.5. This threshold is not fixed and the process of dynamically adjusting the warning threshold will periodically recalculate the mean and standard deviation of risk scores in the recent period and update the threshold When the fluctuation range of real-time operational data increases, as evidenced by a significant recent increase in the standard deviation of the real-time risk score, the system will automatically reduce the coefficient. The value of , thereby reducing the dynamic warning threshold. This makes the early warning mechanism more sensitive during periods of system instability. The comparison between the real-time risk score and the dynamic warning threshold is performed by the logic judgment unit in the monitoring software. Once the score exceeds the threshold at a certain moment, the logic unit immediately sends a Boolean trigger signal to the alarm management module.

[0044] Risk warning signals are generated according to a clear classification system. These signals are categorized into different levels based on the magnitude and duration of the real-time risk score exceeding the dynamic warning threshold, such as "Attention," "Warning," and "Danger." Upon triggering a risk warning signal, the alarm management module activates preset alarm actions. These actions include displaying the alarm information in a prominent color on the monitoring center's human-machine interface, activating the audible and visual alarms, and sending SMS messages and emails containing the device number, alarm time, risk score, and level of exceedance to the mobile terminals of relevant maintenance personnel. All warning signals, trigger times, and related data snapshots are recorded in the alarm log database for post-event analysis and model improvement. Based on the risk warning signal, a corresponding risk mitigation strategy is generated. This strategy is based on a preset strategy knowledge base. The strategy knowledge base is a rule base that stores the mapping relationships between various alarm conditions and recommended actions. These mapping relationships exist in the form of "if-then" rules, such as, "If the real-time risk score exceeds the danger threshold and the vibration risk coefficient contributes more than 70%, then the strategy generated is: It is recommended to immediately implement load reduction operation and arrange for personnel to inspect the compressor bearing status on-site." The risk mitigation strategy is specific and actionable. It includes not only operational suggestions but also brief descriptions of operating procedures, a list of components requiring focused inspection, and links to relevant technical documents. The generated strategy is displayed on the operator's workstation along with the warning information and pushed to mobile devices, providing direct decision support for field personnel. For centrifugal compressors, a typical risk mitigation strategy for high vibration risk might include specific steps such as "reducing the compressor load to 80% of its rated value," "checking the quality and pressure of the lubricating oil," and "rechecking the bearing vibration spectrum using a portable vibration analyzer." The system may also suggest different response times based on the urgency of the risk; for example, for a "warning" level, it recommends handling within 4 hours; for a "hazard" level, immediate action is required.

[0045] The adaptive adjustment mechanism of the dynamic warning threshold is the embodiment of the system intelligence. The adjustment of the dynamic warning threshold is not only based on the standard deviation of the risk score itself, but also introduces other indicators. The system monitors the fluctuation of each original parameter in the real-time running data, such as the fluctuation of the inlet pipe pressure and the fluctuation of the cooling water temperature. If the process parameters fluctuate intensively, even if the standard deviation of the risk score does not change much, the system may moderately reduce the coefficient to improve the early warning sensitivity in advance. Conversely, when the equipment is in an extremely stable working condition, all parameter fluctuations are small, and the system may appropriately increase the coefficient to avoid unnecessary slight fluctuations triggering the warning and reduce false positives. This adjustment enables the dynamic warning threshold to better adapt to different running stages of the equipment and external environmental conditions, achieving intelligent warning. The calculation of the real-time risk score and the warning process constitute a complete monitoring closed loop, with data flow, calculation logic, and response actions closely linked. The monitoring system continuously collects data, the model continuously calculates the score, the threshold is dynamically updated, and once an anomaly is found, an alarm is immediately given and response guidelines are provided. This mechanism transforms offline and periodic risk assessment into online and continuous safety monitoring, greatly enhancing the early detection and warning capability of potential equipment failures. Taking a centrifugal compressor as an example, this implementation can effectively capture early signs of rotor early rub, poor bearing lubrication, and cooling efficiency decline, etc. Through timely risk mitigation strategies, it avoids the expansion of faults and ensures the continuous and stable operation of the production device. The entire process embodies the predictive maintenance concept based on data-driven models and real-time calculation, transforming traditional passive response maintenance into an advanced mode of active early warning and intervention.

[0046] It should be noted that, in this text, relational terms such as first and second are used merely to distinguish one entity or action from another, and do not necessarily require or imply that there is any such actual relationship or order between these entities or actions. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device.

[0047] Although embodiments of the present application 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 therein without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A risk assessment method based on statistical model optimization, characterized in that, The method comprises the following steps: acquiring a running state data set of a target object, the running state data set containing time series data of multiple dimension monitoring indicators; inputting the running state data set into a pre-trained benchmark risk assessment model to generate an initial risk score and a risk distribution feature; iteratively adjusting parameters of the benchmark risk assessment model according to the risk distribution feature through a dynamic correction algorithm to generate an optimized risk assessment model; re-evaluating the running state data set by using the optimized risk assessment model and outputting a corrected risk score and a key risk area identification.

2. The method of risk assessment based on statistical model optimization according to claim 1, characterized in that, The method further comprises the following steps: collecting running parameters of the target object within a preset time range, the running parameters including mechanical vibration data, temperature change data and load fluctuation data; performing spectral decomposition processing on the mechanical vibration data to extract characteristic frequency components and corresponding amplitude change curves; aligning the temperature change data and the load fluctuation data according to timestamps to generate a temperature-load coupling sequence.

3. The method of risk assessment based on statistical model optimization according to claim 2, characterized in that, The method further comprises the following steps: inputting the characteristic frequency components into a first analysis layer of the benchmark risk assessment model to calculate a vibration risk coefficient through a frequency domain feature matching algorithm; inputting the temperature-load coupling sequence into a second analysis layer of the benchmark risk assessment model to generate a thermal risk coefficient through nonlinear regression analysis; fusing the vibration risk coefficient and the thermal risk coefficient to generate an initial risk score and a corresponding risk distribution feature.

4. The method of risk assessment based on statistical model optimization of claim 3, wherein, The method further comprises the following steps: extracting an abnormal fluctuation interval in the risk distribution feature to calculate a duration and a fluctuation amplitude of the abnormal fluctuation interval; determining a correction weight of the benchmark risk assessment model according to a ratio of the duration to the fluctuation amplitude; updating gradient parameters of an output layer of the benchmark risk assessment model based on the correction weight.

5. The method of risk assessment based on optimization of statistical models according to claim 4, characterized in that, The method further comprises the following steps: inputting the characteristic frequency components into a first analysis layer of the optimized risk assessment model to calculate an updated vibration risk coefficient through a corrected frequency domain feature matching algorithm; inputting the temperature-load coupling sequence into a second analysis layer of the optimized risk assessment model to generate an updated thermal risk coefficient through adjusted nonlinear regression analysis; fusing the updated vibration risk coefficient and the updated thermal risk coefficient to output a corrected risk score and a key risk area identification.

6. The method of risk assessment based on statistical model optimization according to claim 5, characterized in that, The method further comprises the following steps: extracting a high-frequency abnormal component in the updated vibration risk coefficient to mark as a mechanical vibration high-risk area; extracting a temperature mutation interval in the updated thermal risk coefficient to mark as a thermal impact high-risk area; superimposing and analyzing spatial distribution data of the mechanical vibration high-risk area and the thermal impact high-risk area to generate the key risk area identification.

7. The method of risk assessment based on optimization of statistical models according to claim 6, characterized in that, The method further comprises the following steps: Obtaining historical maintenance records of the target object, extracting failure occurrence time and failure type in the historical maintenance records; Matching the failure occurrence time with the time stamp of the key risk area identification, calculating the failure prediction accuracy rate; According to the failure prediction accuracy rate, the parameters of the optimized risk assessment model are secondarily corrected.

8. The risk assessment method based on statistical model optimization according to claim 7, characterized in that, The secondary correction of the parameters of the optimized risk assessment model includes: If the failure prediction accuracy rate is lower than the preset threshold, the calculation weight of the vibration risk coefficient is increased; If the failure prediction accuracy rate is higher than the preset threshold, the calculation weight of the thermal risk coefficient is reduced; Based on the adjusted weight coefficient, the corrected risk score and the key risk area identification are regenerated.

9. The risk assessment method based on statistical model optimization of claim 8, wherein, The method further includes: Collecting real-time running data of the target object, inputting the real-time running data into the secondarily corrected risk assessment model, and generating a real-time risk score; When the real-time risk score exceeds the dynamic warning threshold, a risk warning signal is triggered; According to the risk warning signal, a corresponding risk mitigation strategy is generated.

10. The method of risk assessment based on statistical model optimization of claim 9, wherein, The determination process of the dynamic warning threshold includes: Statistically analyzing the distribution of the risk score in the historical running data, calculating the mean and standard deviation of the risk score; According to the linear combination of the mean and standard deviation, the warning threshold is dynamically adjusted; When the fluctuation amplitude of the real-time running data increases, the dynamic warning threshold is reduced.

Images