Automatic instrument fault prediction method based on fuzzy logic

Abstract

The invention discloses an automatic instrument fault prediction method based on fuzzy logic, and belongs to the technical field of industrial automation and equipment fault diagnosis. Comprising the following steps: generating a preprocessed sensor data set; constructing a feature space fusing multi-dimensional information; constructing a three-layer fuzzy reasoning architecture comprising an input layer, a reasoning layer and an output layer; mapping membership function parameters and rule weights in the three-layer fuzzy reasoning architecture into a parameter optimization space comprising a parameter adjustment action, a rule weight optimization action and a reasoning structure adjustment action; self-adaptive adjustment of the membership function parameters is achieved; and calculating the fault probability according to the current optimized membership function parameter, and evaluating the optimization effect of the current parameter configuration in combination with the multi-target reward function. According to the method, through combination of multi-time window dynamic monitoring, multi-type membership function design and a strict hard-soft constraint mechanism, accurate evaluation and stable prediction of the equipment fault probability are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial automation and equipment fault diagnosis, and more particularly to an automation instrument fault prediction method based on fuzzy logic. BACKGROUND

[0002] In modern industrial automation systems, various types of automation instruments are widely used in real-time monitoring and control of key operating states such as temperature, pressure, flow, vibration, and electrical parameters. With the complexity of process flow and the improvement of equipment integration, the operating stability and fault prediction capability of automation instruments are of great significance to ensure the continuous and safe operation of the system. However, instrument equipment is often affected by factors such as temperature changes, mechanical vibrations, electromagnetic interference, environmental humidity, and aging during long-term service, which can cause problems such as response lag, data drift, sudden abnormalities, or functional failure, thereby affecting the accuracy and reliability of process control.

[0003] Existing instrument fault handling methods are mostly post-maintenance or rely on regular maintenance based on human experience. This approach not only has a lagging response, but also has unreasonable maintenance resource allocation, with the risk of misjudgment and missed judgment. Some methods attempt to set fixed thresholds to alarm and judge monitoring data, but it is difficult to adapt to the large fluctuations in working conditions and complex data distribution in the field environment, and the prediction effect is limited. At the same time, although diagnostic techniques based on statistical analysis, neural networks, and other intelligent methods have developed in recent years, these methods still struggle to achieve high accuracy and stability in fault prediction when faced with a large amount of uncertain information, nonlinear change patterns, and multi-source heterogeneous data in industrial sites.

[0004] Fuzzy logic reasoning methods have shown certain advantages in complex system state recognition and trend analysis due to their ability to handle fuzziness, uncertainty, and expert knowledge integration. However, in specific applications, how to design membership function structures, optimize fuzzy rule systems, and improve the adaptability and prediction robustness of the model in a multi-time, multi-parameter environment based on the characteristics of industrial instrument data remains a challenge in engineering practice. In addition, industrial instrument data exhibit both short-term disturbances and long-term degradation, and single-time-scale analysis methods cannot fully reflect the changing trend of equipment health status.

[0005] In summary, how to establish a fuzzy logic prediction mechanism with reasonable structure, responsive sensitivity, and self-adaptive capability based on multi-source sensor information to achieve early prediction and accurate identification of automation instrument faults in complex working conditions has become a technical problem that needs to be solved. SUMMARY

[0006] In order to overcome a series of defects existing in the prior art, the purpose of the present application is to provide an automatic instrument fault prediction method based on fuzzy logic, which comprises the following steps.

[0007] Obtain multi-dimensional sensor raw data in the running process of industrial automation instruments, and perform filtering and denoising, outlier detection and data standardization preprocessing on the multi-dimensional sensor raw data to generate a preprocessed sensor data set.

[0008] Perform dimensionality reduction processing on the preprocessed sensor data set, extract key feature vectors reflecting device health status, running conditions and historical trends, and construct a feature space that integrates multi-dimensional information.

[0009] Construct a three-layer fuzzy reasoning architecture including an input layer, an inference layer and an output layer, wherein the input layer converts continuous key feature vectors into fuzzy set inputs through a membership function, the inference layer performs inference calculation based on fuzzy rules, and the output layer outputs device fault probability evaluation results.

[0010] Map the membership function parameters and rule weights in the three-layer fuzzy reasoning architecture to a parameter optimization space containing parameter adjustment actions, rule weight optimization actions and reasoning structure adjustment actions.

[0011] Construct a multi-objective reward function containing a false alarm rate minimization objective, a missed alarm rate minimization objective and a stability maintenance objective, combine the membership function monotonicity hard constraint, the rule weight normalization hard constraint, the reasoning architecture integrity hard constraint and the performance stability soft constraint, and realize adaptive adjustment of the membership function parameters.

[0012] During the adaptive adjustment process, a multi-time window monitoring mechanism is established to monitor the change trend of the key feature vectors in real time, the fault probability is calculated according to the current optimized membership function parameters, and the optimization effect of the current parameter configuration is evaluated combined with the multi-objective reward function.

[0013] If the optimization effect does not meet the preset performance indicators or the number of continuous iterations exceeds the preset threshold, continue to adjust the membership function parameters based on the feedback signal of the multi-objective reward function until the performance indicator requirements are met or the maximum number of iterations is reached.

[0014] If the optimization effect has reached the preset performance indicators, fix the current membership function parameter configuration, integrate the prediction results of different time scales based on the multi-time window monitoring mechanism, and output a comprehensive diagnostic report containing fault type identification, occurrence probability evaluation, predicted occurrence time and result confidence.

[0015] Further, the acquisition process of the multi-dimensional sensor raw data includes: through the multi-type sensor nodes deployed on site, real-time collection of equipment running state parameters according to a preset sampling frequency, wherein the temperature data sampling frequency is set to 0.1-1Hz to adapt to the thermal inertia characteristics; the pressure and flow data sampling frequency is set to 1-10Hz to capture the process dynamics; the vibration data sampling frequency is set to 1000-10000Hz to meet the Nyquist sampling theorem requirements; the current data sampling frequency is set to 50-500Hz to detect electrical fault characteristics; a sensor data timestamp synchronization mechanism based on network time protocol is established, through master clock synchronization and clock drift compensation, the microsecond-level accurate alignment of different types of sensor data in the time dimension is ensured, and through triple-mode redundant sensor configuration and data integrity verification based on parity check, fault tolerance processing of single-point sensor is realized, when detecting abnormality of a certain sensor node, according to the voting logic, the backup sensor is automatically switched or the adjacent sensor data is interpolated and compensated through the weighted least squares method, ensuring the continuity and reliability of multi-dimensional sensor data acquisition.

[0016] Further, the input layer in the three-layer fuzzy inference architecture adopts a multi-type membership function design, including the following steps: for continuous characteristic variables, a combination of trapezoidal membership function, Gaussian membership function and triangular membership function is adopted, wherein the trapezoidal membership function is used to represent the fault state interval with clear boundaries, the Gaussian membership function is used to model the health state parameters with continuous change characteristics, and the triangular membership function is used to describe the transition state and critical working condition; for discrete characteristic variables, single-point membership function and interval membership function are adopted, the parameter range of each membership function is determined through expert knowledge and historical data statistics; according to the newly collected sensor data and fault label information, the function parameters are dynamically adjusted by using the optimization strategy combining gradient descent method and genetic algorithm, including the four vertex coordinates of trapezoidal function, the mean and variance parameters of Gaussian function, and the vertex position of triangular function, to ensure that the membership function can accurately reflect the fuzzy characteristics and distribution law of characteristic variables under different running states.

[0017] Further, the parameter adjustment action space covers continuous adjustment of membership function shape parameters, center position parameters and boundary width parameters, the adjustment range and direction of each parameter are represented by real number coding method, and the adjustment range is limited within the reasonable interval determined by expert knowledge and physical constraints; the rule weight optimization action space includes single rule weight adjustment, rule combination weight batch adjustment and adaptive weight distribution based on performance feedback, and the sum of all rule weights is kept as 1 through normalization constraint; the inference structure adjustment action space covers dynamic adjustment of rule base structure, including adding new rules, deleting redundant rules, refining rule conditions and modifying rule conclusions.

[0018] Further, the false positive rate minimization objective is quantified by calculating the proportion of samples that are predicted to be faulty but are actually normal, and a negative reward mechanism is designed to penalize excessively high false positive rates; the false negative rate minimization objective is quantified by calculating the proportion of samples that are actually faulty but are predicted to be normal; the stability preservation objective is evaluated by calculating the variance and rate of change of the prediction results within consecutive time windows, and a stability reward mechanism is designed to encourage the smoothness and consistency of the prediction results; a Pareto optimal solution set search method is used to find the best balance point between false positive rate, false negative rate and stability, and a multi-objective genetic algorithm or multi-objective particle swarm algorithm is used to generate multiple candidate solutions, and the most suitable parameter configuration scheme is selected according to the actual application requirements.

[0019] Further, the monotonicity hard constraint of the membership function is verified by mathematics to ensure that the boundary slope of the trapezoidal function and the triangular function meets the monotonicity requirement, and the symmetry and continuity of the Gaussian function meet the mathematical definition; the rule weight normalization hard constraint is ensured by a real-time monitoring and automatic adjustment mechanism to ensure that the sum of the weights of all active rules is always equal to 1, and when the weight adjustment causes the sum to deviate from 1, automatic scaling correction is performed; the reasoning architecture integrity hard constraint verifies the connection integrity between the input layer, reasoning layer and output layer of the fuzzy reasoning architecture, ensuring that each input variable has a corresponding membership function and each rule has a clear antecedent and consequent definition; the performance stability soft constraint sets a prediction accuracy decline threshold and a performance fluctuation tolerance range, and gives different degrees of penalty signal when the performance deviates from the expected range, guiding the parameter optimization process to converge towards a stable performance; when a hard constraint violation is detected, the current parameter adjustment is immediately stopped and rolled back to the last valid configuration, and when a soft constraint violation is detected, the violation degree is recorded and the corresponding penalty is increased in subsequent optimization.

[0020] Further, the multi-time window includes three levels of short-term window, medium-term window and long-term window, wherein: the short-term window length is set to 5-10 sampling periods, mainly for capturing rapidly changing fault signs; the medium-term window length is set to 50-100 sampling periods, for monitoring the slow degradation process of device performance; the long-term window length is set to 500-1000 sampling periods, for analyzing seasonal changes and long-term trends in device operation; each time window is updated in a sliding window manner, and the window overlap rate is set to 50% to ensure continuity of information, and within each time window, statistical quantities including mean, standard deviation, maximum, minimum and change trend are calculated, and by comparing the changes in statistical quantities in different time windows, different stages of fault development are identified.

[0021] Further, the calculation of the failure probability comprises the following steps: calculating the fuzzy set membership value corresponding to each input characteristic variable according to the current optimized membership function parameter, and calculating the activation strength and output contribution of each rule; based on the conditional independence assumption between fault types established by statistical analysis of historical failure data, calculating the posterior probability of each fault type under the current characteristic condition through a Bayesian network; combining the historical probability prediction sequence in the sliding time window and the current prediction result, calculating the time-fused failure probability through an exponentially weighted moving average, and balancing the real-time and stability of the prediction result.

[0022] Further, the fault type identification provides the top 3-5 fault types and their corresponding identification confidence through multi-classification probability output and confidence sorting, and simultaneously gives the typical feature description and possible fault cause analysis of each fault type; the occurrence probability evaluation displays the fault occurrence probability in a combination of percentage and grade classification, and is accompanied by corresponding color coding and graphical display; the predicted occurrence time provides the time window of fault occurrence, including the interval estimation of the earliest occurrence time, the most likely occurrence time and the latest occurrence time, based on the equipment degradation model and fault development rate prediction; the result confidence provides the reliability evaluation for the entire diagnosis result by comprehensively considering the data quality, model uncertainty and historical prediction accuracy factors.

[0023] Compared with the prior art, the present application has the following beneficial effects.

[0024] The present application realizes accurate evaluation and stable prediction of the equipment failure probability by combining multi-time window dynamic monitoring, multi-type membership function design and strict hard-soft constraint mechanism. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 A flowchart of a fuzzy logic-based automatic instrument fault prediction method disclosed in an embodiment of the present application. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme in the embodiments of the present application will be described in more detail below with reference to the drawings in the embodiments of the present application. In the drawings, the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The described embodiments are part of the embodiments of the present application, but not all the embodiments.

[0027] Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the scope of protection of the present application.

[0028] The embodiments described below and the terms used therein are examples, and are intended to explain the present application, and should not be interpreted as limiting the present application.

[0029] As Figure 1 shown, an automatic instrument fault prediction method based on fuzzy logic includes the following steps.

[0030] Step 1, obtaining multi-dimensional sensor original data in the running process of industrial automation instruments, and performing filtering and denoising, outlier detection and data standardization preprocessing on the multi-dimensional sensor original data to generate a preprocessed sensor data set.

[0031] Step 2, performing dimensionality reduction processing on the preprocessed sensor data set, extracting key feature vectors reflecting device health status, running conditions and historical trends, and constructing a feature space integrating multi-dimensional information.

[0032] Step 3, constructing a three-layer fuzzy reasoning architecture including an input layer, an inference layer and an output layer, wherein the input layer converts continuous key feature vectors into fuzzy set inputs through a membership function, the inference layer performs inference calculation based on fuzzy rules, and the output layer outputs device fault probability evaluation results.

[0033] Step 4, mapping the membership function parameters and rule weights in the three-layer fuzzy reasoning architecture to a parameter optimization space containing parameter adjustment actions, rule weight optimization actions and reasoning structure adjustment actions.

[0034] Step 5, constructing a multi-objective reward function including a false alarm rate minimization objective, a missed alarm rate minimization objective and a stability maintenance objective, combining the membership function monotonicity hard constraint, the rule weight normalization hard constraint, the reasoning architecture integrity hard constraint and the performance stability soft constraint, and realizing adaptive adjustment of the membership function parameters.

[0035] Step 6, in the adaptive adjustment process, a multi-time window monitoring mechanism is established to monitor the change trend of the key feature vectors in real time, the fault probability is calculated according to the current optimized membership function parameters, and the optimization effect of the current parameter configuration is evaluated in combination with the multi-objective reward function.

[0036] Step 7, if the optimization effect does not meet the preset performance indicators or the number of continuous iterations exceeds the preset threshold, the membership function parameters are continuously adjusted based on the feedback signal of the multi-objective reward function until the performance indicator requirements are met or the maximum number of iterations is reached.

[0037] Step 8, if the optimization effect has reached the preset performance indicators, the current membership function parameter configuration is fixed, the prediction results of different time scales are integrated based on the multi-time window monitoring mechanism, and a comprehensive diagnostic report containing fault type identification, occurrence probability evaluation, predicted occurrence time, and result confidence is output.

[0038] In step 1, the multi-dimensional data covers temperature, pressure, current, voltage, vibration, and other types, covering all-around state information of instrument equipment operation. However, the data collected by sensors often has noise, interference, and outliers, which directly affects the accuracy of subsequent fault diagnosis. Therefore, filtering and denoising operation can effectively eliminate random noise and periodic interference, improve the signal-to-noise ratio of data; outlier detection avoids misjudgment caused by device sudden abnormality or sensor failure; data standardization preprocessing eliminates the dimension difference of different sensor data, making it comparable and consistent. Through these steps, the final preprocessed data set not only has high quality, but also more truly reflects the running state of the equipment. This step ensures the quality of input data for subsequent model training and inference, avoids the problem of "garbage in, garbage out", and is the key guarantee to improve fault prediction accuracy and stability.

[0039] Step 2 focuses on dimensionality reduction of preprocessed multi-dimensional sensor data, aiming to extract key feature vectors that can effectively reflect device health status, operating conditions, and historical trends from a large amount of redundant or weakly correlated data. Dimensionality reduction not only reduces the computational complexity of data, reduces resource consumption of model training and inference, but also avoids the "curse of dimensionality" problem caused by high dimensionality, which is beneficial to model generalization. Through scientific design of feature extraction methods (such as principal component analysis, linear discriminant analysis, automatic encoder, etc.), multi-dimensional sensor information can be organically integrated to build a compact and representative feature space. This feature space integrates the static properties and dynamic trends of the device, improving the sensitivity to potential fault states. This step abstracts and compresses key features, improving the efficiency and accuracy of subsequent fuzzy logic reasoning, while reducing the interference of redundant information on model judgment, ensuring that the prediction model still has strong robustness and adaptability in complex and variable industrial environments.

[0040] Step 3 constructs a three-layer fuzzy reasoning system composed of input layer, inference layer, and output layer, which becomes the core logic structure for fault probability assessment. The input layer converts continuous key feature vectors into fuzzy sets by designing membership functions, breaking through the limitations of traditional hard threshold classification and allowing inputs to be expressed in a fuzzy and uncertain form, which enhances the adaptability to industrial site uncertainties and complexities. The inference layer conducts reasoning based on fuzzy rules constructed from expert experience and historical data, which can flexibly handle complex and variable causal relationships, make comprehensive judgments on input information, and improve the accuracy and interpretability of fault identification. The output layer converts the fuzzy reasoning results into specific fault probability assessment. This architecture, through the nonlinear mapping of fuzzy logic and the coordination of multiple rules, not only has high sensitivity and accuracy, but also can make more humanized and explanatory fault prediction for complex industrial equipment operating conditions.

[0041] Step 4 innovatively maps the membership function parameters and rule weights in the fuzzy reasoning architecture into a parameter optimization space containing parameter adjustment actions, rule weight optimization actions, and reasoning structure adjustment actions, thereby achieving fine-tuned control of the fuzzy reasoning process. By establishing an optimization space, the boundaries, shapes of fuzzy sets, and relative importance of rules can be dynamically adjusted, breaking through the limitations of traditional fuzzy reasoning relying on fixed rules and parameters. The parameter optimization space enables the reasoning model to have self-adaptive ability, automatically adjusting the form of fuzzy rules and membership functions for different equipment and operating conditions, and realizing personalized customization of the fault prediction model. This step avoids overfitting and underfitting, ensuring high-performance fault prediction in diverse and complex industrial environments.

[0042] Step 5 realizes adaptive optimization and adjustment of membership function parameters by designing a multi-objective reward function containing minimum false alarm rate, minimum missed alarm rate, and stability maintenance, combined with membership function monotonicity hard constraints, rule weight normalization hard constraints, reasoning architecture integrity hard constraints, and performance stability soft constraints. The multi-objective reward function considers the accuracy and stability of the fault prediction model, balances the contradiction between false alarms and missed alarms, and avoids the bias caused by single-target optimization. Hard constraints ensure the reasonableness and logical consistency of model parameters, preventing ineffective or abnormal configurations during the tuning process. Soft constraints promote smooth changes in model performance and stable performance over a long period of operation.

[0043] Step 6 designs a multi-time window monitoring mechanism to achieve real-time monitoring of the changing trend of key feature vectors at different time scales, combined with the current optimized membership function parameter calculation device failure probability, and based on the multi-objective reward function to evaluate the optimization effect of the current parameter configuration. Through the design of multi-time window, the diversity of information of short-term fluctuations and long-term trends of the device can be captured, and the timeliness and accuracy of fault detection can be improved. The real-time evaluation mechanism ensures the dynamic feedback of the parameter optimization process, which can timely discover and correct the optimization deviation, and avoid falling into local optimum. This mechanism enables the entire prediction process to form a closed-loop adaptive adjustment capability, continuously improving the accuracy and stability of fault prediction, meeting the dual demands of real-time and accuracy in complex industrial environments.

[0044] Step 7 ensures the effective convergence of the optimization process and the controllability of the model performance. By setting reasonable performance indicators and iteration thresholds, it avoids ineffective over-tuning, saving computing resources and time costs. At the same time, through continuous feedback adjustment, the model can gradually approach the optimal state in a complex and variable environment, improving the stability and accuracy of fault prediction.

[0045] Step 8 realizes the intelligent integration of multi-dimensional and cross-period data, improving the comprehensiveness and accuracy of the prediction results. Fault type identification provides a clear direction for device maintenance, probability of occurrence quantifies the risk level, predicted occurrence time assists in early warning, and result confidence provides a reliable reference for decision-making, greatly improving the scientificity and initiative of device management.

[0046] In summary, the automatic instrument fault prediction method based on fuzzy logic systematically realizes the intelligent management of the whole process from data collection, preprocessing, feature extraction, to fuzzy reasoning architecture construction, parameter optimization, multi-objective adaptive adjustment, and result integration. Its advantage lies in fully combining multi-dimensional data fusion and fuzzy logic nonlinear reasoning, breaking through the limitations of traditional fault prediction relying on hard rules and static models, and giving strong adaptive learning and real-time optimization capabilities. The multi-objective reward function and multi-time window monitoring mechanism effectively balance the accuracy, stability, and real-time requirements, ensuring the reliability and practicality of the prediction results. This method not only improves the accuracy and efficiency of automatic instrument fault prediction, but also enhances the adaptability and robustness to complex industrial environments, greatly improving the safety of device operation and the initiative of maintenance.

[0047] Further, the acquisition process of the multi-dimensional sensor raw data includes: through the multi-type sensor nodes deployed on site, real-time collection of equipment operation state parameters according to a preset sampling frequency, wherein the temperature data sampling frequency is set to 0.1-1 Hz to adapt to the thermal inertia characteristics; the pressure and flow data sampling frequency is set to 1-10 Hz to capture the process dynamics; the vibration data sampling frequency is set to 1000-10000 Hz to meet the Nyquist sampling theorem requirement; the current data sampling frequency is set to 50-500 Hz to detect electrical fault characteristics; a sensor data timestamp synchronization mechanism based on network time protocol is established, through master clock synchronization and clock drift compensation, the microsecond-level accurate alignment of different types of sensor data in the time dimension is ensured, and through triple modular redundancy sensor configuration and data integrity check based on parity check, fault tolerance processing of single-point sensor is realized, when detecting abnormality of a certain sensor node, according to the voting logic, the backup sensor is automatically switched or the adjacent sensor data is interpolated and compensated through the weighted least squares method, ensuring the continuity and reliability of multi-dimensional sensor data acquisition.

[0048] In summary, by differentiating the sampling frequency of various sensors to match the dynamic characteristics of different physical quantities, the data collection is ensured to be both timely and representative; at the same time, the network time protocol synchronization mechanism is introduced, effectively solving the alignment problem of multi-source data on the time axis, avoiding the interference of asynchronous errors on subsequent analysis; combined with triple modular redundancy configuration and data integrity check means, the detection and repair capability of sensor faults is enhanced, the robustness and fault tolerance of the data acquisition link are improved, thereby providing continuous, accurate and reliable multi-dimensional data support for subsequent fault prediction models.

[0049] Further, the input layer in the three-layer fuzzy reasoning architecture adopts a multi-type membership function design, including the following steps: for continuous feature variables, a combination of trapezoidal membership function, Gaussian membership function and triangular membership function is adopted, wherein the trapezoidal membership function is used to represent the fault state interval with clear boundaries, the Gaussian membership function is used to model the health state parameters with continuous change characteristics, and the triangular membership function is used to describe the transition state and critical working condition; for discrete feature variables, single-point membership function and interval membership function are adopted, the parameter range of each membership function is determined through expert knowledge and historical data statistics; according to the newly collected sensor data and fault label information, the gradient descent method and genetic algorithm are combined to dynamically adjust the function parameters, including the four vertex coordinates of the trapezoidal function, the mean and variance parameters of the Gaussian function, and the vertex position of the triangular function, to ensure that the membership function can accurately reflect the fuzzy characteristics and distribution law of feature variables under different operating conditions.

[0050] In summary, by introducing multiple types of membership function combinations, different types of characteristic variables are modeled differently, effectively improving the expression ability of the input layer to diversified operating states. The trapezoidal, Gaussian, and triangular functions are used to describe the clear boundaries, gradual trends, and critical transitions of continuous variables, improving the resolution of complex state intervals. The single-point and interval membership functions set by expert knowledge assist in realizing accurate fuzzy mapping of discrete fault signals. Combined with gradient descent and genetic algorithm for dynamic optimization of membership function parameters, the model has the ability to adapt to new data and state changes, enhancing the accuracy and flexibility of the reasoning process, ensuring that fuzzy inputs can accurately map real working condition characteristics, providing a stable and reliable input basis for downstream fuzzy rule reasoning.

[0051] Further, the fuzzy rule construction of the reasoning layer adopts a multi-source knowledge fusion strategy, including the following steps: collecting the fault diagnosis experience of field experts, converting the experience knowledge into "IF-THEN" form fuzzy rules, and establishing an initial fuzzy rule base; based on historical fault data, mining potential rule patterns, automatically extracting the association rules between fault symptoms and fault types from a large number of historical operation data; identifying redundant rules and conflicting rules through rule similarity calculation and semantic analysis methods, and solving rule conflict problems using weight voting and confidence sorting strategies; according to the feedback information of new fault cases and prediction results, continuously expanding and optimizing the fuzzy rule base through incremental learning, and regularly cleaning up inefficient rules and obsolete rules to maintain the simplicity and effectiveness of the rule base.

[0052] In summary, by fusing expert knowledge and data-driven means, a fuzzy rule system with experience and adaptability is constructed. Expert experience is converted into formal "IF-THEN" rules, providing an initial diagnostic framework, while patterns automatically mined based on historical data significantly improve the coverage and update speed of rules. By introducing rule similarity analysis and semantic conflict identification mechanisms, redundant and conflicting rules are effectively eliminated, ensuring the logical consistency and reasoning efficiency of the rule reasoning process. At the same time, with the help of weight voting and confidence sorting strategies, the decision-making influence of key rules is strengthened. The incremental learning mechanism realizes the continuous adaptation and dynamic update of the rule base to new fault scenarios, while the regular cleaning strategy maintains the simplicity and efficiency of the rule set.

[0053] Further, the output layer failure probability evaluation result calculation adopts a multi-level fusion mechanism, including the following steps: calculating the activation degree of each fuzzy rule through T-norm operator and S-norm operator, wherein the T-norm operator adopts minimum value operator, algebraic product operator or Lukasiewicz product operator, and the S-norm operator adopts maximum value operator, algebraic sum operator or Lukasiewicz sum operator; assigning a dynamic weight to each rule according to the historical accuracy, confidence and support degree of the rule, and ensuring that the sum of all rule weights is 1 through weight normalization processing; for multiple possible failure types, calculating the occurrence probability of each failure type through weighted average, maximum value selection or fuzzy integral, and simultaneously calculating the overall failure risk score; quantifying the uncertainty of the prediction result through interval mathematics and possibility theory method, and outputting the probability evaluation result containing confidence interval and credibility level.

[0054] In summary, through the multi-level fusion mechanism, accurate calculation and uncertainty quantification of failure probability are realized, which significantly improves the prediction credibility and practical value of the model. Through the combination of T-norm and S-norm operators in multiple forms, the activation calculation requirements of different types of fuzzy rules are effectively adapted, enhancing the model's expression ability for rule logic; the introduction of dynamic weight allocation strategy based on historical performance enables high-reliability rules to play a greater role in reasoning, while weight normalization processing ensures the stability and interpretability of the overall reasoning process. In failure probability inference, methods such as weighted average, maximum value selection and fuzzy integral are flexibly adopted to ensure that the model can accurately reflect the risk level of multiple potential failure types; combining interval mathematics and possibility theory, the uncertainty of the prediction result is measured, so that the output not only includes the probability value, but also provides confidence interval and credibility level.

[0055] Further, the parameter adjustment action space covers continuous adjustment of membership function shape parameters, center position parameters and boundary width parameters, and the adjustment range is limited within a reasonable interval determined by expert knowledge and physical constraints; the rule weight optimization action space includes single rule weight adjustment, rule combination weight batch adjustment and adaptive weight allocation based on performance feedback, and normalization constraint is used to ensure that the sum of all rule weights remains 1; the reasoning structure adjustment action space covers dynamic adjustment of rule base structure, including adding new rules, deleting redundant rules, refining rule conditions and modifying rule conclusions.

[0056] In summary, by constructing a multi-dimensional parameter adjustment action space covering the membership function parameters, rule weights, and reasoning structure, the overall dynamic optimization of the fuzzy reasoning model is achieved. The shape, position, and width parameters of the membership function are encoded as real numbers, combined with expert knowledge and physical constraints to limit the adjustment range, ensuring that parameter updates are both flexible and consistent with actual working conditions; the rule weight optimization not only supports fine-grained adjustment of individual weights, but also covers batch adjustment of combined weights, and through the normalization mechanism, the stability of the weight sum is maintained, enhancing the overall coordination of the model; the reasoning structure adjustment allows dynamic maintenance of the rule base, including adding new rules, removing redundancies, and refining conditions and conclusions, improving the accuracy and adaptability of the rule system.

[0057] Further, the false alarm rate minimization objective is quantified by calculating the proportion of samples predicted as faulty but actually normal, and a negative reward mechanism is designed to punish high false alarm rates; the false negative rate minimization objective is quantified by calculating the proportion of samples that are actually faulty but predicted as normal; the stability maintenance objective is evaluated by calculating the variance and change rate of the prediction results within a continuous time window, and a stability reward mechanism is designed to encourage the smoothness and consistency of the prediction results; a Pareto optimal solution set search method is used to find the best balance point between false alarm rate, false negative rate, and stability, and a multi-objective genetic algorithm or multi-objective particle swarm algorithm is used to generate multiple candidate solutions, and the most suitable parameter configuration scheme is selected according to the actual application requirements.

[0058] In summary, by constructing the quantification objectives of false alarm rate, false negative rate, and prediction stability, the overall optimization of the fault prediction model performance is achieved. False alarm rate and false negative rate reflect the shortcomings of the model in false alarms and missed faults, respectively, and the negative reward mechanism designed for false alarms effectively suppresses excessive alarms, improving the reliability of the prediction; the false negative rate index ensures that fault risks are not ignored, ensuring the timeliness of safety warnings; the stability objective analyzes the volatility of the prediction results within a continuous time window, promoting the smoothness and consistency of the model output, which helps to avoid the decision-making confusion caused by the dramatic fluctuations of the prediction results. A Pareto optimal solution set search method combined with a multi-objective evolutionary algorithm is used to systematically explore the trade-off relationship between multiple objectives and generate diverse parameter configuration schemes.

[0059] Further, the membership function monotonicity hard constraint is verified by mathematics to ensure that the boundary slope of the trapezoidal function and the triangular function meets the monotonicity requirement, and the symmetry and continuity of the Gaussian function meet the mathematical definition; the rule weight normalization hard constraint is ensured by real-time monitoring and automatic adjustment mechanism to ensure that the sum of the weights of all active rules is always equal to 1, and when the weight adjustment causes the sum to deviate from 1, automatic scaling correction is performed; the reasoning architecture integrity hard constraint verifies the connection integrity between the input layer, the reasoning layer and the output layer of the fuzzy reasoning architecture, ensuring that each input variable has a corresponding membership function, and each rule has a clear antecedent and consequent definition; the performance stability soft constraint sets a prediction accuracy decline threshold and a performance fluctuation tolerance range, and when the performance deviates from the expected range, different degrees of penalty signals are given to guide the parameter optimization process to converge to the direction of performance stability; when the hard constraint violation is detected, the current parameter adjustment is immediately stopped and rolled back to the nearest valid configuration, and when the soft constraint violation is detected, the violation degree is recorded and the corresponding penalty strength is increased in the subsequent optimization.

[0060] In summary, the multi-level constraint mechanism ensures the structural rationality and performance stability of the fuzzy reasoning model. The monotonicity hard constraint of the membership function uses mathematical verification method to strictly regulate the morphological characteristics of the trapezoidal, triangular and Gaussian functions, ensuring the continuity and logical consistency of the input mapping process; the rule weight normalization hard constraint monitors the weight sum in real time, automatically performs proportional correction, and maintains the normativity and coordination of the rule weight distribution; the reasoning architecture integrity hard constraint ensures the tightness of the connection between the input, reasoning and output layers, preventing logical omission or error transmission; the performance stability soft constraint sets a performance fluctuation threshold and applies a penalty to guide the parameter optimization to smooth convergence. When the hard constraint is triggered, it is immediately rolled back to avoid risks caused by invalid adjustments, and when the soft constraint is violated, the penalty strength is dynamically adjusted. The dual protection mechanism effectively maintains the efficient and stable operation of the model, improves the accuracy and reliability of fault prediction.

[0061] Further, the multi-time window includes three levels of short-term window, medium-term window and long-term window, wherein: the short-term window length is set to 5-10 sampling periods, mainly used to capture rapidly changing fault signs; the medium-term window length is set to 50-100 sampling periods, used to monitor the slow degradation process of device performance; the long-term window length is set to 500-1000 sampling periods, used to analyze the seasonal changes and long-term trends of device operation; each time window is updated in a sliding window manner, and the window overlap rate is set to 50% to ensure the continuity of information, and the feature statistics including mean, standard deviation, maximum, minimum and change trend are calculated in each time window, and the different stages of fault development are identified by comparing the changes of the statistics in different time windows.

[0062] In summary, by setting three levels of multi-time windows, namely short-term, medium-term and long-term, the dynamic monitoring of equipment state changes is achieved. The short-term window focuses on capturing instantaneous and rapid fault signals, improving the sensitivity to sudden abnormalities; the medium-term window focuses on the gradual decline of equipment performance, which can timely discover potential degradation trends and support preventive maintenance decisions; the long-term window reveals periodic influences and structural changes by analyzing seasonal fluctuations and long-term operation trends of the equipment. The use of sliding windows and high overlap rate design ensures the continuity and integrity of the monitoring data, avoiding information omission and faults, and the multi-dimensional calculation of statistical quantities enhances the identification ability of different fault stages. As can be seen from the above, the combination of multi-time scales significantly improves the timeliness and accuracy of fault detection, and realizes the dynamic and detailed description of equipment state.

[0063] Further, the calculation of the fault probability comprises the following steps: calculating the fuzzy set membership value corresponding to each input feature variable according to the current optimized membership function parameter, and calculating the activation strength and output contribution of each rule; based on the statistical analysis of historical fault data, the conditional independence assumption between fault types is established, and the posterior probability of each fault type under the current feature condition is calculated by Bayesian network; combining the historical probability prediction sequence and the current prediction result in the sliding time window, the fault probability after time fusion is calculated by exponential weighted moving average, balancing the real-time and stability of the prediction result.

[0064] In summary, by fusing fuzzy membership calculation and Bayesian network probability inference, accurate probability estimation of fault types under multi-feature input conditions is effectively realized. The optimization of membership function parameters ensures the accuracy of fuzzy set mapping of input data, and further improves the rationality of rule activation strength and output contribution; the conditional independence assumption constructed by historical data enables the Bayesian network to scientifically calculate the posterior probability of each fault type, strengthening the theoretical basis of probability inference; combined with sliding time window and exponential weighted moving average method, dynamic fusion of historical and real-time prediction information is realized, which not only ensures the timeliness of prediction, but also improves the smoothness and stability of output results. As can be seen from the above, the application of multi-level fusion strategy improves the accuracy and robustness of fault probability calculation, providing more reliable and continuous support for equipment state monitoring and fault warning.

[0065] Further, the performance indicators include a prediction accuracy indicator, a prediction precision indicator, a prediction recall rate indicator, a prediction stability indicator, and a response time indicator, wherein: the prediction accuracy indicator is quantified by calculating the proportion of correctly predicted fault and normal state samples in the total samples; the prediction precision indicator is evaluated by calculating the mean square error or the mean absolute error between the predicted fault probability and the actual fault occurrence; the prediction recall rate indicator is measured by calculating the proportion of successfully predicted fault samples in the actual fault samples; the prediction stability indicator is evaluated by calculating the variance and change rate of consecutive prediction results, requiring the prediction results to remain basically consistent under similar operating conditions; and the response time indicator requires that the processing time from data input to output of the prediction result does not exceed the preset real-time requirement.

[0066] In summary, through the design of multi-dimensional performance indicators, comprehensive evaluation of the overall capability of the fault prediction system is achieved. The prediction accuracy indicator reflects the correctness of the overall judgment of the model, ensuring reasonable differentiation between fault and normal state; the prediction precision indicator quantifies the error between the prediction probability and the actual fault, improving the delicacy of fault risk assessment; the prediction recall rate indicator emphasizes the completeness of fault identification, ensuring that critical abnormalities are not missed; the prediction stability indicator analyzes the volatility of the prediction results, ensuring the continuity and reliability of the model output under similar working conditions; and the response time indicator limits the system processing delay, meeting the strict real-time requirements of industrial sites.

[0067] Further, the fault type identification is provided through multi-class probability output and confidence sorting, providing the top 3-5 fault types and their corresponding identification confidence, as well as typical feature descriptions and possible fault cause analysis for each fault type; the occurrence probability evaluation adopts a combination of percentage and grade classification to display the fault occurrence probability, with corresponding color coding and graphical display; the predicted occurrence time is based on the equipment degradation model and fault development rate prediction, providing a time window for fault occurrence, including interval estimation of the earliest occurrence time, the most likely occurrence time, and the latest occurrence time; and the result confidence provides a reliability evaluation for the entire diagnosis result by considering factors such as data quality, model uncertainty, and historical prediction accuracy.

[0068] In summary, by combining multi-classification probability output with confidence ranking, the multiple possibilities and their relative importance of potential device failures are accurately presented, improving the transparency and interpretability of failure identification. By providing typical feature descriptions and failure cause analysis, users can better understand the prediction results and support targeted maintenance decisions. The occurrence probability evaluation uses a combination of percentages and level classification, supplemented by intuitive color coding and graphical display, enhancing the readability and interactivity of the information. The predicted occurrence time is based on a dynamic device degradation model, accurately determining the time interval for potential failure, facilitating the advance arrangement of maintenance resources. The result confidence evaluation considers multiple factors to ensure the reliability and scientificity of the diagnosis report.

[0069] Finally, it should be pointed out that the above examples are only used to illustrate the technical solutions of the present application, and are not limiting. Although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can modify the technical solutions described in the foregoing examples, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting faults in automated instruments based on fuzzy logic, characterized in that, Includes the following steps: The system acquires raw multidimensional sensor data during the operation of industrial automation instruments, and performs filtering, noise reduction, outlier detection, and data standardization preprocessing on the raw multidimensional sensor data to generate a preprocessed sensor dataset. The preprocessed sensor dataset is subjected to dimensionality reduction processing to extract key feature vectors reflecting the equipment's health status, operating conditions, and historical trends, and a feature space integrating multi-dimensional information is constructed. A three-layer fuzzy inference architecture is constructed, comprising an input layer, an inference layer, and an output layer. The input layer converts continuous key feature vectors into fuzzy set inputs through a membership function. The inference layer performs inference calculations based on fuzzy rules. The output layer outputs the device fault probability assessment results. The membership function parameters and rule weights in the three-layer fuzzy inference architecture are mapped into a parameter optimization space that includes parameter adjustment actions, rule weight optimization actions, and inference structure adjustment actions; A multi-objective reward function is constructed, comprising objectives of minimizing false positive rate, minimizing false negative rate, and maintaining stability. This function is then combined with hard constraints on the monotonicity of the membership function, hard constraints on the normalization of rule weights, hard constraints on the integrity of the inference architecture, and soft constraints on performance stability to achieve adaptive adjustment of the membership function parameters. During the adaptive adjustment process, a multi-time-window monitoring mechanism is established to monitor the changing trend of the key feature vectors in real time, calculate the fault probability based on the currently optimized membership function parameters, and evaluate the optimization effect of the current parameter configuration in conjunction with the multi-objective reward function. If the optimization effect fails to meet the preset performance index or the number of consecutive iterations exceeds the preset threshold, the membership function parameters will continue to be adjusted based on the feedback signal of the multi-objective reward function until the performance index requirement is met or the maximum number of iterations is reached. If the optimization effect has reached the preset performance index, then the current membership function parameter configuration is fixed, and the prediction results of different time scales are integrated based on the multi-time window monitoring mechanism to output a comprehensive diagnostic report including fault type identification, occurrence probability assessment, expected occurrence time and result confidence.

2. The method for predicting faults in automated instruments based on fuzzy logic according to claim 1, characterized in that, The process of acquiring raw data from the multi-dimensional sensors includes: real-time acquisition of equipment operating status parameters through multiple types of sensor nodes deployed on-site, according to a preset sampling frequency. Specifically, the temperature data sampling frequency is set to 0.1–1 Hz to accommodate thermal inertia; the pressure and flow data sampling frequency is set to 1–10 Hz to capture process dynamics; the vibration data sampling frequency is set to 1000–10000 Hz to meet the Nyquist sampling theorem requirements; and the current data sampling frequency is set to 50–500 Hz to detect electrical fault characteristics. A sensor data timestamp synchronization mechanism based on a network time protocol is established. Master clock synchronization and clock drift compensation ensure microsecond-level precise alignment of different types of sensor data in the time dimension. Triple modulo redundancy sensor configuration and parity-based data integrity verification enable fault-tolerant handling of single-point sensor failures. When an anomaly is detected in a sensor node, automatic switching to a backup sensor is performed according to voting logic, or adjacent sensor data interpolation compensation is performed using weighted least squares method to ensure the continuity and reliability of multi-dimensional sensor data acquisition.

3. The method for predicting faults in automated instruments based on fuzzy logic according to claim 1, characterized in that, The input layer of the three-layer fuzzy inference architecture employs a multi-type membership function design, including the following steps: For continuous feature variables, a combination of trapezoidal, Gaussian, and triangular membership functions is used. The trapezoidal membership function represents fault state intervals with clear boundaries, the Gaussian membership function models health state parameters with continuously changing characteristics, and the triangular membership function describes transitional states and critical conditions. For discrete feature variables, single-point and interval membership functions are used, with the parameter range of each membership function determined through expert knowledge and historical data statistics. Based on newly acquired sensor data and fault tag information, an optimization strategy combining gradient descent and genetic algorithms is used to dynamically adjust the function parameters, including the coordinates of the four vertices of the trapezoidal function, the mean and variance parameters of the Gaussian function, and the vertex positions of the triangular function, ensuring that the membership functions accurately reflect the fuzzy characteristics and distribution patterns of feature variables under different operating conditions.

4. The method for predicting faults in automated instruments based on fuzzy logic according to claim 1, characterized in that, The parameter adjustment action space encompasses the continuous adjustment of membership function shape parameters, center position parameters, and boundary width parameters. The adjustment magnitude and direction of each parameter are represented by real number encoding, and the adjustment range is limited to a reasonable interval determined by expert knowledge and physical constraints. The rule weight optimization action space includes the adjustment of individual rule weights, the batch adjustment of rule combination weights, and adaptive weight allocation based on performance feedback. Normalization constraints ensure that the sum of all rule weights remains at 1. The inference structure adjustment action space covers the dynamic adjustment of the rule base structure, including the addition of new rules, the deletion of redundant rules, the refinement of rule conditions, and the correction of rule conclusions.

5. The method for predicting faults in automated instruments based on fuzzy logic according to claim 1, characterized in that, The goal of minimizing the false positive rate is quantified by calculating the proportion of samples predicted as faulty but actually normal, and a negative reward mechanism is designed to penalize excessively high false positive rates. The goal of minimizing the false negative rate is quantified by calculating the proportion of samples that are actually faulty but predicted as normal. The goal of maintaining stability is evaluated by calculating the variance and rate of change of the prediction results within a continuous time window, and a stability reward mechanism is designed to encourage the smoothness and consistency of the prediction results. The Pareto optimal solution set search method is used to find the best balance between false positive rate, false negative rate and stability. Multiple candidate solutions are generated by a multi-objective genetic algorithm or a multi-objective particle swarm optimization algorithm, and the most suitable parameter configuration scheme is selected according to the actual application requirements.

6. The method for predicting faults in automated instruments based on fuzzy logic according to claim 1, characterized in that, The membership function monotonicity hard constraint is mathematically verified to ensure that the boundary slopes of trapezoidal and triangular functions meet the monotonicity requirements, and that the symmetry and continuity of the Gaussian function satisfy the mathematical definition. The rule weight normalization hard constraint ensures that the sum of the weights of all active rules is always equal to 1 through real-time monitoring and automatic adjustment mechanisms. When weight adjustment causes the sum to deviate from 1, automatic scaling correction is performed. The inference architecture integrity hard constraint verifies the connection integrity between the input, inference, and output layers of the fuzzy inference architecture, ensuring that each input variable has a corresponding membership function and each rule has a clear antecedent and consequent definition. The performance stability soft constraint sets a threshold for prediction accuracy decline and a performance fluctuation tolerance range. When performance deviates from the expected range, different levels of penalty signals are given to guide the parameter optimization process to converge towards performance stability. When a hard constraint violation is detected, the current parameter adjustment is immediately stopped and rolled back to the most recent effective configuration. When a soft constraint violation is detected, the degree of violation is recorded and the corresponding penalty is increased in subsequent optimizations.

7. The method for predicting faults in automated instruments based on fuzzy logic according to claim 1, characterized in that, The multi-time window system comprises three levels: short-term, medium-term, and long-term windows. The short-term window, with a length of 5–10 sampling periods, is primarily used to capture rapidly changing fault symptoms. The medium-term window, with a length of 50–100 sampling periods, is used to monitor the slow degradation process of equipment performance. The long-term window, with a length of 500–1000 sampling periods, is used to analyze seasonal variations and long-term trends in equipment operation. Each time window is updated using a sliding window method, with a window overlap rate of 50% to ensure information continuity. Characteristic statistics, including mean, standard deviation, maximum, minimum, and trend, are calculated within each time window. By comparing the changes in these statistics across different time windows, different stages of fault development are identified.

8. The method for predicting faults in automated instruments based on fuzzy logic according to claim 1, characterized in that, The calculation of the fault probability includes the following steps: calculating the fuzzy set membership value corresponding to each input feature variable based on the currently optimized membership function parameters, and calculating the activation intensity and output contribution of each rule; establishing the conditional independence assumption between fault types based on statistical analysis of historical fault data, and calculating the posterior probability of each fault type under the current feature conditions through a Bayesian network; combining the historical probability prediction sequence within the sliding time window and the current prediction result, calculating the fault probability after time fusion through exponential weighted moving average, balancing the real-time performance and stability of the prediction result.

9. The method for predicting faults in automated instruments based on fuzzy logic according to claim 1, characterized in that, The fault type identification provides the top 3-5 fault types and their corresponding identification confidence levels through multi-class probability output and confidence ranking, while also giving typical feature descriptions and possible fault cause analyses for each fault type. The occurrence probability assessment displays the fault occurrence probability using a combination of percentage and grade classification, accompanied by corresponding color coding and graphical display. The predicted occurrence time is based on the equipment degradation model and fault development rate prediction, providing a time window for the fault occurrence, including interval estimates of the earliest occurrence time, the most likely occurrence time, and the latest occurrence time. The result confidence level provides a credibility assessment for the entire diagnostic result by comprehensively considering data quality, model uncertainty, and historical prediction accuracy.

Images