A trend failure prediction method based on dynamic mode and threshold cooperation

Abstract

The present application relates to the technical field of industrial equipment state monitoring and fault diagnosis, in particular to a trend fault prediction method based on dynamic mode and threshold cooperation. In the present application, a nonlinear mapping between working conditions and key parameters is established to generate a theoretical prediction interval that dynamically changes with load in real time, effectively eliminating parameter drift interference caused by working condition fluctuations, combining maintenance records to quantify the real-time health baseline of the equipment, adjusting the warning threshold width and narrowness according to the feature similarity and health level, realizing adaptive monitoring of different aging stages in the whole life cycle, using a multi-dimensional feature vector to fuse physical field information for pattern matching, comprehensively evaluating the deviation severity and shape similarity through fuzzy reasoning, and locking the anomaly in advance according to the high feature coincidence degree when the numerical value is not seriously out of limit, accurately capturing early weak signs and outputting the abnormal source, significantly improving the diagnosis accuracy under variable working conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial equipment state monitoring and fault diagnosis, and in particular to a trend fault prediction method based on dynamic mode and threshold cooperation. BACKGROUND

[0002] The technical field of industrial equipment state monitoring and fault diagnosis mainly covers the continuous acquisition, storage and analysis of operating parameters such as vibration, temperature, pressure, current, voltage, speed and the like generated by equipment such as power plant steam turbine generators, boilers, pumps, fans and transformers during long-term operation, and the identification and determination of equipment degradation process, fault evolution characteristics and abnormal behavior by establishing a description method of equipment operating state change over time.

[0003] Among them, the trend fault prediction method refers to a method that, for the performance degradation and hidden faults gradually generated by key equipment in the operation process of the power plant, forms a trend curve by sorting the historical operation data by time, based on the long-term changes of specific operating indicators such as vibration amplitude, temperature rise rate, pressure offset, current fluctuation amplitude, etc., and uses methods such as setting threshold intervals, slope change points and cumulative offsets to interpret trend characteristics, thereby making an early determination of possible fault states of the equipment. It usually completes trend recognition by comparing and analyzing data at the same measuring point in different operating cycles.

[0004] The prior art only constructs a trend curve according to time sorting and relies on fixed thresholds to make longitudinal comparison of the same measuring point. This static logic ignores the dynamic shaping of parameter theoretical benchmarks by complex and variable working conditions, resulting in normal drift during load regulation being mistaken for fault characteristics. Relying on single-dimensional linear extrapolation cannot decouple environmental disturbances and equipment responses, making it difficult to distinguish between condition-induced fluctuations and performance degradation. The rigid threshold lacks adaptability to the health evolution throughout the life cycle and cannot adaptively adjust the boundary according to the degree of aging. In non-steady state operation, it is easy to miss weak signs or produce false positives due to benchmark lag. SUMMARY

[0005] The purpose of the present application is to solve the shortcomings in the prior art and to propose a trend fault prediction method based on dynamic mode and threshold cooperation.

[0006] In order to achieve the above-mentioned purpose, the present application adopts the following technical scheme, a trend fault prediction method based on dynamic mode and threshold cooperation, comprising the following steps:

[0007] S1: Collecting key operating parameters and current working condition parameters of target power plant equipment in a specified continuous operating cycle and classifying and processing to obtain equipment operating state data sequence;

[0008] S2: determine the reference data section in the device operation state data sequence, estimate the predicted value of the key operation parameter at the future time point according to the reference data section, calculate the predicted deviation value of the predicted value and the corresponding current value, and select the parameter trend prediction data;

[0009] S3: based on the device operation state data sequence, a plurality of feature vectors of the target power plant equipment are constructed, compared with the stored fault mode feature set, and the feature similarity score is calculated to obtain the fault feature matching result;

[0010] S4: the current working condition parameter in the device operation state data sequence is combined and regressed, the reference floating curve of each current working condition parameter under multiple working conditions is fitted by constructing a back propagation neural network, and a dynamic early warning judgment reference is determined;

[0011] S5: calculate the deviation severity index of the predicted deviation value and the dynamic early warning judgment reference, and combine the feature similarity score for early warning state judgment to obtain the equipment fault early warning and diagnosis result.

[0012] As a further scheme of the application, the device operation state data sequence includes vibration intensity data, current fluctuation data, temperature rise amplitude data, fluid pressure data, load rate, environmental temperature data, cumulative running time and main shaft speed data arranged in time sequence, the parameter trend prediction data includes predicted deviation value and marked predicted offset indicator, the fault feature matching result includes feature similarity score and updated existing feature set, the dynamic early warning judgment reference includes threshold baseline generated based on reference floating curve and threshold adjustment ratio, and the equipment fault early warning and diagnosis result includes early warning state judgment and determined abnormal source.

[0013] As a further scheme of the application, the acquisition step of the device operation state data sequence is specifically:

[0014] S111: collect vibration intensity, current fluctuation, temperature rise amplitude and fluid pressure data of the target power plant equipment in a specified continuous running period to form key operation parameters, construct time series data corresponding to the key operation parameters, compare and identify abnormal extreme values of the key operation parameters with the preset physical limit threshold, remove the abnormal extreme values in the time series data, and generate key operation parameter time series;

[0015] S112: obtain the load rate of the target power plant equipment, the temperature data of the running environment of the target power plant equipment, the cumulative running time of the target power plant equipment and the main shaft speed data of the target power plant equipment, form the current working condition parameter, map the current working condition parameter to a unified time dimension, process the current working condition parameters with inconsistent dimensions, and establish the current working condition parameter set.

[0016] S113: According to the current working condition parameter set, the operation condition categories of various power plant equipment are divided, each key operation parameter in the key operation parameter time sequence is classified into the corresponding operation condition category, and device operation state data sequences are obtained.

[0017] As a further scheme of the present application, the parameter trend prediction data acquisition step is specifically:

[0018] S211: The latest recorded data of a specified duration in the device operation state data sequence is intercepted to form a real-time monitoring data segment, the numerical difference between the real-time monitoring data segment and the data segment of the same duration located in the specified time interval in the device operation state data sequence is calculated, and the data segment with a numerical difference less than a preset matching threshold is determined as the reference data section;

[0019] S212: Based on the numerical change of the reference data section, the predicted value of each key operation parameter at future time points is estimated, the correction compensation amount of the predicted value is calculated, and the predicted value is numerically corrected to generate a corrected key operation parameter predicted value;

[0020] S213: The difference between the current value of the key operation parameter in the device operation state data sequence and the corresponding corrected key operation parameter predicted value is calculated to obtain a prediction deviation value, and the parameter value of the prediction deviation value exceeding the 95% numerical distribution interval in the device operation state data sequence is marked as a prediction offset indicator. The prediction deviation value and the prediction offset indicator are integrated to generate parameter trend prediction data.

[0021] As a further scheme of the present application, the fault feature matching result acquisition step is specifically:

[0022] S311: Based on the vibration intensity in the device operation state data sequence, the vibration amplitude change rate of the target power plant equipment is calculated, based on the current fluctuation in the device operation state data sequence, the current harmonic offset value of the target power plant equipment is calculated, and based on the temperature rise amplitude in the device operation state data sequence, the temperature rise gradient data of the target power plant equipment is calculated. The vibration amplitude change rate, the current harmonic offset value and the temperature rise gradient data are constructed into a feature vector to obtain a multi-dimensional operation feature vector;

[0023] S312: The stored fault mode feature set is acquired, the multi-dimensional operation feature vector and the stored fault mode feature set are normalized, the cosine similarity between the normalized multi-dimensional operation feature vector and the normalized stored fault mode feature set is calculated, the quantification score of feature matching is obtained based on the cosine similarity, and a feature similarity score is generated;

[0024] S313: judging whether the current state matches the known failure mode in the stored failure mode feature set according to the feature similarity score, identifying new potential failure mode features from the device running state data sequence and updating the stored failure mode feature set to obtain a failure feature matching result.

[0025] As a further scheme of the present application, the acquisition step of the dynamic early warning judgment reference is specifically:

[0026] S411: performing multivariate regression analysis calculation on the current working condition parameter set in the device running state data sequence, fitting the reference floating curve of each current working condition parameter under multiple working conditions by constructing a back propagation neural network, inputting the current working condition parameter set into the reference floating curve, calculating the key operation parameter theoretical prediction interval, and taking the key operation parameter theoretical prediction interval as the threshold baseline of the target power plant equipment under the current running state;

[0027] S412: collecting the maintenance record of the target power plant equipment, quantifying the overall health level of the target power plant equipment in combination with the cumulative running length of the target power plant equipment, and establishing the current health baseline of the equipment;

[0028] S413: calculating the threshold adjustment ratio required to trigger early warning in the threshold baseline according to the feature similarity score and the current health baseline of the equipment, and performing adjustment on the threshold baseline by using the threshold adjustment ratio to obtain the dynamic early warning judgment reference.

[0029] As a further scheme of the present application, the acquisition step of the device failure early warning and diagnosis result is specifically:

[0030] S511: filtering the prediction deviation value marked by the prediction offset indicator from the parameter trend prediction data, calculating the ratio of the filtered prediction deviation value to the adjusted threshold baseline boundary value in the dynamic early warning judgment reference to obtain a deviation severity index, and normalizing and combining the deviation severity index with the feature similarity score to construct a decision input vector;

[0031] S512: taking a preset failure state label as a label input, constructing a relationship weight mapping between the decision input vector and the label input, and using a fuzzy classification algorithm to perform early warning state judgment on the decision input vector according to the relationship weight mapping to output an early warning state judgment result;

[0032] S513: When the pre-warning state determination result is triggering pre-warning, the bias severity index and the feature similarity score in the decision input vector are numerically sorted, the object with the largest value is determined as the dominant feature, the dominant abnormal source is determined according to the dominant feature, the fault type is determined in combination with the fault feature matching result, and the device fault pre-warning and diagnosis result is output.

[0033] Compared with the prior art, the advantages and positive effects of the present application are that:

[0034] In the present application, the theoretical prediction interval dynamically changing with load is generated in real time by establishing the nonlinear mapping between the working condition and the key parameters, the parameter drift interference caused by working condition fluctuation is effectively eliminated, the real-time health baseline of the equipment is quantified in combination with the maintenance record, the warning threshold width is cooperatively adjusted according to the feature similarity and the health level, the adaptive monitoring of different aging stages in the whole life cycle is realized, the physical field information is fused by using the multi-dimensional feature vector to perform pattern matching, the bias severity and the shape similarity are comprehensively evaluated by fuzzy reasoning, the abnormality is locked in advance according to the high feature coincidence degree when the numerical value is not seriously over-limit, the early weak signs are accurately captured and the abnormal source is output, and the diagnosis accuracy under variable working conditions is significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The present application is a trend fault prediction method based on dynamic mode and threshold cooperation, and the flow chart of the overall architecture is shown in the figure;

[0036] Figure 2 The present application is a device running state data sequence construction flow chart;

[0037] Figure 3 The present application is a parameter trend prediction data generation flow chart;

[0038] Figure 4 The present application is a fault feature matching and mode updating flow chart;

[0039] Figure 5 The present application is a dynamic pre-warning determination reference determination flow chart;

[0040] Figure 6 The present application is a device fault pre-warning and diagnosis decision flow chart. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application, and are not used to limit the present application.

[0042] Please refer to Figure 1The application provides a technical scheme, a trend fault prediction method based on dynamic mode and threshold cooperation, comprising the following steps:

[0043] S1: collect key operation parameters and current working condition parameter of the target power plant equipment in a specified continuous operation period, and perform classification processing to obtain an equipment operation state data sequence;

[0044] S2: determine a reference data section in the equipment operation state data sequence, estimate a predicted value of the key operation parameter at a future time point according to the reference data section, calculate a prediction deviation value of the predicted value and a corresponding current value, and select parameter trend prediction data;

[0045] S3: construct a plurality of feature vectors of the target power plant equipment based on the equipment operation state data sequence, compare the feature vectors with a stored fault mode feature set, calculate a feature similarity score, and obtain a fault feature matching result;

[0046] S4: perform combined regression analysis on the current working condition parameter in the equipment operation state data sequence, fit a reference floating curve of each current working condition parameter under a plurality of working conditions by constructing a back propagation neural network, and determine a dynamic early warning judgment reference;

[0047] S5: calculate a deviation severity index of the prediction deviation value and the dynamic early warning judgment reference, normalize and combine the feature similarity score to construct a decision input vector, perform early warning state judgment, and obtain an equipment fault early warning and diagnosis result.

[0048] The equipment operation state data sequence comprises vibration intensity data, current fluctuation data, temperature rise amplitude data, fluid pressure data, load rate, environmental temperature data, cumulative operation time length and main shaft speed data arranged in time sequence, the parameter trend prediction data comprises a prediction deviation value and a marked prediction offset indicator, the fault feature matching result comprises a feature similarity score and an updated existing feature set, the dynamic early warning judgment reference comprises a threshold baseline generated based on the reference floating curve and a threshold adjustment proportion, and the equipment fault early warning and diagnosis result comprises early warning state judgment and a determined abnormal source.

[0049] Please refer to Figure 2 , the acquisition step of the equipment operation state data sequence is specifically as follows:

[0050] S111: collect vibration intensity, current fluctuation, temperature rise amplitude and fluid pressure data of the target power plant equipment in a specified continuous operation period to form key operation parameters, construct time sequence data corresponding to the key operation parameters, compare and identify abnormal extreme values of the key operation parameters with a preset physical limit threshold value, remove the abnormal extreme values in the time sequence data, and generate key operation parameter time sequence;

[0051] Firstly, the analog signals are acquired in parallel through the sensor array deployed at the key positions of the pump body, in which the piezoelectric vibration acceleration sensor captures the bearing vibration, the Hall current sensor monitors the motor stator current, the platinum resistance measures the winding temperature, and the pressure transmitter collects the outlet pressure. All analog signals are converted into digital signals by a high-precision 24-bit analog-to-digital converter in the form of independent sampling per channel, and are uniformly stamped with a global positioning system time stamp accurate to milliseconds by a field programmable gate array chip, forming four initial time series. Subsequently, abnormal extreme value identification and elimination based on statistical principles are performed. The setting of the physical limit threshold is first based on the high-pressure feed water pump in the past year full operating condition history database, and 10,000 vibration intensity sampling points in the full load stable running state are filtered out to form a sample set. Through statistical calculation, the arithmetic mean of the set (derived from the total sample set divided by the number of samples) is 3.2 mm / s, and the sample standard deviation (derived from the arithmetic square root of the sample variance) is 0.8 mm / s. According to the normal distribution three standard deviation criterion, the decision interval is set, and the upper limit of the physical limit threshold is calculated, the calculation formula is , that is, 5.6 mm / s is set as the upper limit threshold of the valid data. The setting value judgment possibility contains two kinds: if the monitored value is less than or equal to 5.6, it is judged as valid physical data; if the monitored value is greater than 5.6, it is judged as non-physical noise. When traversing the real-time vibration intensity time series collected, the current detected value 12.4 mm / s is obtained, and the value comparison logic is performed. Since , the value falls into the possibility interval of non-physical noise, it is judged that the value deviates from the possible motion range of the physical entity, and belongs to the random pulse noise caused by strong electromagnetic interference, so it is eliminated from the sequence. At the same time, the manual control instructions are retrieved from the historical operation log database of the power plant central control system through the industrial standard communication interface, and the time stamp interval of 09:00:00 to 09:05:00 is matched through the database query statement, and it is identified that the operator issued the "load adjustment" instruction during this period. Therefore, the 5-minute time period is marked as a non-steady state transition interval. Finally, the vibration, current, temperature and pressure data in the noise points and non-steady state intervals identified above are executed in batch elimination operation. For the time breakpoints generated after elimination, linear interpolation is used to connect the small gaps with a length of less than 1 second, and the breakpoints with a length greater than 1 second are retained as segmentation markers, and finally the time series of key operating parameters of the device only containing the true physical characteristics under automatic stable control are generated.

[0052] S112: Obtain the load rate of the target power plant equipment, the temperature data of the running environment of the target power plant equipment, the cumulative running time of the target power plant equipment, and the main shaft speed data of the target power plant equipment, form the current working condition parameter, map the current working condition parameter to a unified time dimension, associate the current working condition parameters with different dimensions, and establish the current working condition parameter set.

[0053] Firstly, the real-time working condition data associated with the high-pressure feed water pump is read through the industrial Ethernet communication interface at a frequency of 10 times per second. The data sources include the real-time load rate of the motor in the distributed control system (the read value is 85%), the environmental temperature uploaded by the plant environmental monitor (the read value is 25 degrees Celsius), the cumulative running time in the equipment full life cycle management database (the read value is 12000 hours), and the main shaft speed feedback by the key phase sensor (the read value is 2980 revolutions per minute). In view of the time alignment problem caused by the inconsistent sampling frequencies of different source data, linear mapping and synchronization processing based on time stamp are performed. Taking the time stamp 10:00:01 in the time sequence of the key operating parameters as the reference anchor point, for the environmental temperature data with a sampling frequency of 1 time per minute, the nearest two valid collection values before and after this time are retrieved, that is, the value 25.0 degrees Celsius collected at the previous time 10:00:00 and the value 25.2 degrees Celsius collected at the next time 10:00:02, and linear interpolation logic is used for calculation. The calculation formula is That is, the environmental temperature mapped at this time is 25.1 degrees Celsius. Similarly, the load rate and the cumulative running time are interpolated and mapped at the millisecond level. Subsequently, the current working condition parameter set is established. The load rate 85%, the environmental temperature 25.1 degrees Celsius, the cumulative running time 12000 hours and the main shaft speed 2980 revolutions per minute at the same time point (10:00:01) are combined to construct a four-dimensional feature vector as the working condition background description at this time, and the data integrity checking algorithm is used to check whether there is a null value or a non-numeric identifier in the vector. After confirming that there is no missing value, the current working condition parameter set corresponding to the key operating parameter time sequence is finally established.

[0054] S113: According to the current working condition parameter set, the running working condition categories of various power plant equipment are divided, each key operating parameter in the key operating parameter time sequence is classified into the corresponding running working condition category, and the equipment running state data sequence is obtained;

[0055] First, according to the load rate and the spindle speed in the current working condition parameter set, the two core variables are clustered and divided, and the preset working condition category judgment rule is based on the iterative clustering analysis results of the factory performance curve and historical operation data of the water pump. By analyzing the distribution of historical operation data for one year, the data space is divided into three possible discrete interval categories: the first category is "high load full load working condition", the judgment interval is that the load rate is greater than 90% and the speed is greater than 2950 revolutions per minute; the second category is "medium load regulation working condition", the judgment interval is that the load rate is between 60% and 90% and the speed is greater than 2900 revolutions per minute; the third category is "low load standby working condition", the judgment interval is that the load rate is less than 60%. When performing specific grouping operations, read the working condition parameters at the current time (load rate 85%, speed 2980 revolutions per minute), and through logical comparison operation: the value 85% is in the closed interval of 60% and 90%, and the value The combination falls within the definition range of the second category, so the key operating parameters at this time (such as vibration intensity 3.5 mm / s, current fluctuation 120 A, etc.) are attached with the metadata label of "medium load regulation working condition". Finally, traverse the entire time sequence, and use the hash mapping table to quickly distribute all data points to the corresponding working condition category storage buckets according to the above rules, for example, extract all vibration, current, temperature data marked as "medium load regulation working condition" time and reorganize them in memory, link them in chronological order, and form a device operating state data sequence that specifically describes the device performance under this specific working condition.

[0056] Please refer to Figure 3 The parameter trend prediction data acquisition step is specifically:

[0057] S211: Cut the latest recorded data of a specified time length in the device operating state data sequence to form a real-time monitoring data segment, calculate the numerical difference between the real-time monitoring data segment and the same length data segment in the device operating state data sequence within the specified time interval, and determine the data segment with a numerical difference less than the preset matching threshold as the reference data segment;

[0058] Firstly, the latest recorded specified duration data is intercepted from the equipment running state data sequence using a first-in-first-out queue, and the specified duration is set to 60 seconds. The vibration intensity data within the past 60 seconds (including 6000 sampling points) is intercepted, and after denoising processing, a real-time monitoring data segment is formed. Then, a sliding window search is performed within the specified time interval (set to the past 30 days of historical storage data). The sliding window length is also set to 60 seconds, and the step length is 1 second. The shape similarity between the real-time monitoring data segment and each historical sliding window segment is calculated. The average Euclidean distance calculation logic is used to quantify the numerical difference. Assuming that the value of the first point of the real-time segment is 3.5, and the value of the corresponding point of the historical segment is 3.4, the values of the second points are 3.6 and 3.5 respectively. The square of the difference value of the corresponding points is calculated, and the sum is divided by the total number of points 6000, and then the square root is taken. For example, the distance value between a certain historical segment and the real-time segment is calculated to be 0.08. The preset matching threshold is set based on the statistical results of the autocorrelation of historical data under the same working condition. The average distance between two adjacent normal running data segments under the "medium load regulation working condition" in history (the statistical value is 0.1) and the standard deviation of the distance (the statistical value is 0.02) are calculated. According to the principle of statistical confidence interval, there are two possibilities in distance judgment: if the distance value is less than 0.14, it is determined to be matched; if the distance value is greater than or equal to 0.14, it is determined to be not matched. The critical value calculation formula is . Finally, the calculated distance value 0.08 is compared with the preset matching threshold 0.14. Since , it is determined that the historical segment matches the possibility, and the waveform shape and numerical level are highly consistent with the current real-time segment. It is determined as the reference data segment.

[0059] S212: Based on the numerical change of the reference data segment, the predicted values of each key running parameter at future time points are estimated, and the correction compensation of the predicted values is calculated and the numerical correction of the predicted values is performed to generate the corrected key running parameter predicted values;

[0060] Firstly, the preliminary estimation is based on the reference data segment. The determined reference data segment is located in the historical database, and the data value of the first second after the reference data segment in the historical time axis (the historical true record value is 3.8 mm / s) is directly used as the preliminary predicted value of the first second in the future. Then, the trend deviation degree is calculated. The least square linear regression calculation is performed on the data of the real-time monitoring data segment within the last 10 seconds and the data of the reference data segment within the corresponding last 10 seconds, respectively. The slope of the real-time data is 0.05, and the slope of the reference data is 0.03. The difference between the two is calculated as the trend difference, and the calculation formula is , which indicates that the current device vibration rate is faster than the historical similar time. Then calculate the correction compensation, introduce the correction coefficient (set to 1.5), which is derived from the backtest experiment statistics of the past 50 fault cases, aiming to give greater weight compensation when the short-term trend diverges. The correction compensation is calculated as the offset degree multiplied by the correction coefficient, and the calculation formula is . Finally, the predicted value is corrected, and the preliminary predicted value 3.8 is added to the correction compensation 0.03, and the calculation formula is , that is, the corrected key operating parameter prediction value of the first second in the future is 3.83 mm / s.

[0061] S213: Calculate the difference between the current value of the key operating parameter in the device operating state data sequence and the corresponding corrected key operating parameter prediction value to obtain the prediction deviation value. Mark the parameter value that exceeds the 95% value distribution interval in the device operating state data sequence as the prediction offset indicator. Integrate the prediction deviation value and the prediction offset indicator to generate parameter trend prediction data.

[0062] First, when the first second in the future actually arrives, the actual measurement value of the key operating parameter at that time (for example, 3.95 mm / s) is collected through the sensor, and the difference between it and the corrected prediction value 3.83 calculated at the last time is calculated, and the calculation formula is , and the obtained 0.12 is the prediction deviation value. Then determine the determination standard of the prediction offset indicator. By statistically analyzing the prediction deviation values of the last 1000 normal time points in the device operating state data sequence, a probability distribution model of the deviation value is fitted, and the average value of the normal deviation is calculated to be 0 and the standard deviation is 0.04. According to the 95% confidence interval theory of normal distribution, there are two possibilities for deviation state: if the deviation value is less than or equal to 0.0784, it belongs to normal fluctuation; if the deviation value is greater than 0.0784, it belongs to abnormal deviation. The calculation formula of the upper limit of the determination interval is . Then compare and mark the prediction deviation value 0.12 calculated at the current time with the upper limit 0.0784 of the distribution interval, since , it indicates that the deviation value falls into the possibility interval of abnormal deviation, and the system immediately tags the parameter value with an electronic label of "prediction offset indicator" and records its deviation degree. Finally, integrate the data to package and encapsulate the prediction deviation value 0.12, the actual measurement value 3.95 and the "prediction offset indicator" label to generate the parameter trend prediction data at that time.

[0063] Please refer to Figure 4 , so the acquisition step of the fault feature matching result is:

[0064] S311: Calculate the vibration amplitude change rate of the target power plant equipment based on the vibration intensity in the equipment operating state data sequence, calculate the current harmonic offset value of the target power plant equipment based on the current fluctuation in the equipment operating state data sequence, calculate the temperature rise gradient data of the target power plant equipment based on the temperature rise amplitude in the equipment operating state data sequence, construct the vibration amplitude change rate, the current harmonic offset value and the temperature rise gradient data into a feature vector to obtain a multi-dimensional operating feature vector;

[0065] First, calculate the vibration amplitude change rate, extract the vibration intensity time domain waveform data in the last 1 second from the cache, find the maximum and minimum values through the extreme value search logic, calculate the difference (peak-to-peak value) as 0.5 millimeters per second, divide by the time interval 1 second (from the sampling setting), the calculation formula is , that is, the vibration amplitude change rate is 0.5. Then calculate the current harmonic offset value, perform frequency spectrum transformation processing on the stator current data of the last one power frequency period (0.02 seconds), extract the amplitude of the 5th harmonic component in the frequency spectrum as 2.5 amperes, query the standard 5th harmonic amplitude reference in the database under the same working condition as 1.0 amperes, calculate the relative offset value, the calculation formula is , that is, the current harmonic offset value is 1.5 (150%). Then calculate the temperature rise gradient data, select the sampling data of the stator winding temperature sensor in the last 10 minutes, calculate the temperature change slope using linear fitting logic, measure that the temperature rises by 2 degrees Celsius, divide by the time interval 10 minutes, the calculation formula is , that is, the temperature rise gradient is 0.2 degrees Celsius per minute. Finally, construct the feature vector, arrange the above three dimensionless or standardized calculation results in a fixed order to form a three-dimensional feature array [0.5, 1.5, 0.2], which is the multi-dimensional operating feature vector describing the current equipment abnormal state.

[0066] S312: Obtain the stored fault mode feature set, normalize the multi-dimensional operating feature vector and the stored fault mode feature set, calculate the cosine similarity between the normalized multi-dimensional operating feature vector and the normalized stored fault mode feature set, obtain the quantitative score of feature matching based on the cosine similarity, and generate a feature similarity score;

[0067] The stored fault mode feature set includes: a standard fault sample vector library and corresponding fault category definition labels cleaned and calibrated from historical operation and maintenance data;

[0068] Each sample vector in the standard fault sample vector library is composed of historical vibration amplitude change rate, historical current harmonic offset value and historical temperature rise gradient data collected under the confirmed fault working condition, arranged and combined in the same dimension order as when constructing the multi-dimensional operating feature vector;

[0069] The fault category definition label includes a vibration anomaly class label representing the degree of mechanical wear of the rotating component, a current anomaly class label representing the electrical insulation state of the stator and rotor, and a temperature rise anomaly class label representing the cooling efficiency of the heat exchange process.

[0070] First, the vector is normalized, and the real-time multi-dimensional running feature vector [0.5, 1.5, 0.2] generated in step S311 is calculated. The Euclidean norm (module length) is calculated as , and each element in the vector is divided by the module length to obtain the unit direction vector [0.31, 0.94, 0.13]. Similarly, the standard fault sample vector of "bearing wear" is retrieved from the fault mode feature library, and the original value is assumed to be [0.6, 1.4, 0.3]. The module length is calculated as , and the normalized standard vector is [0.39, 0.90, 0.19]. Then the cosine similarity is calculated, and the dot product operation is performed on the two normalized vectors, and the calculation formula is . Finally, the score is generated, and the calculation result 0.9916 is the feature similarity score of the current state and the "bearing wear" fault mode. All fault modes in the library (such as rotor imbalance, stator short circuit, etc.) are traversed, and the above calculation is repeated. Assuming that the similarity with other modes is less than 0.5, 0.9916 is selected as the final feature similarity score.

[0071] S313: Determine whether the current state matches the known fault mode in the stored fault mode feature set according to the feature similarity score, identify new potential fault mode features from the device running state data sequence, and update the stored fault mode feature set to obtain a fault feature matching result;

[0072] First, set the matching confirmation threshold, which is set based on statistical analysis of historical diagnosis accuracy. By analyzing the confusion matrix of the past 500 automatic diagnosis results and manual review results, it is determined that the false positive rate is very low when the similarity is higher than 0.85. Therefore, the threshold is set to 0.85. The matching state includes two possibilities: if the score is greater than 0.85, it is determined to be a successful match of the known mode; if the score is less than or equal to 0.85, it is determined to be a match of the known mode. The highest similarity score 0.9916 calculated in step S312 is compared with the threshold value, and since , the current state is determined to fall into the possibility interval of successful matching, and the matching result is directly output as "bearing wear". If the highest similarity score calculated is only 0.45 (less than 0.85), an unsupervised learning mechanism is triggered, a density-based spatial clustering logic is called, the neighborhood radius parameter (derived from experience setting) is set to 0.3, the minimum number of points parameter is set to 5, the current unmatched feature vector is put into the cache area, it is assumed that 10 similar unmatched vectors have been accumulated in the cache area, and the average Euclidean distance between them is calculated to be only 0.2. The density possibility is determined: if the average distance is less than the radius 0.3, it is determined that a high-density cluster is formed; if it is greater than 0.3, it is determined to be discrete noise. Since , the algorithm determines that these points form a new high-density clustering. The centroid of the new cluster is calculated by adding the corresponding dimension values of the 10 vectors and taking the arithmetic mean. Assuming that the calculated centroid vector is [0.8, 0.2, 0.9]. Finally, the set is updated, and the centroid vector [0.8, 0.2, 0.9] is stored as a newly discovered fault mode in the feature set, and a temporary label "potential fault mode_01" is automatically assigned. At this time, the output fault feature matching result is "newly discovered potential fault mode_01".

[0073] Please refer to Figure 5 , the steps for obtaining the dynamic early warning judgment benchmark are as follows:

[0074] S411: Perform multivariate regression analysis on the current working condition parameter set in the device running state data sequence, fit the benchmark floating curve of each current working condition parameter under multiple working conditions by constructing a back propagation neural network, input the current working condition parameter set into the benchmark floating curve, calculate the key running parameter theoretical prediction interval, and take the key running parameter theoretical prediction interval as the threshold baseline of the target power plant equipment under the current running state;

[0075] First, a back propagation neural network is constructed. An error back propagation neural network with a three-layer structure is adopted. The input layer contains 2 neurons, which receive the current working condition parameters (load rate, environmental temperature) respectively. The hidden layer contains two levels, each level is set to 10 neurons, and linear rectification logic is used to increase the nonlinear mapping ability. The output layer contains 4 neurons, which output the theoretical prediction values of vibration intensity, current fluctuation, temperature rise amplitude and fluid pressure, respectively. During model training, the mean square error is used as the loss function, and the gradient descent logic based on the first moment estimation (learning rate is set to 0.001) is used to update the weights based on the historical normal running data set. The current working condition (load rate , environmental temperature normalized value ) is fed as an input vector to the trained model, and forward propagation calculation is performed on the output layer. According to the calculation formula , : No. Theoretical predicted values ​​of key operating parameters ( Corresponding vibration, Corresponding current, Corresponding to the temperature rise, Corresponding pressure); : No. The physical saturation coefficient of each parameter is determined based on the mechanical or electrical physical limits of each parameter of the equipment, and represents the theoretical maximum value that the parameter may reach under extreme physical conditions. The base of the natural logarithm (approximately 2.718); : No. The input condition features for the first The weight of each output parameter reflects the strength of the influence of specific operating conditions (such as load or temperature) on that specific operating parameter. The larger the value, the stronger the driving effect of the change in the input operating condition on the output parameter. : No. Each input condition feature value (i.e., the aforementioned) Load factor and Ambient temperature); : No. The activation bias of each output neuron is used to adjust the activation threshold of this parameter, which determines the position of the curve's translation on the horizontal axis. : Input dimension, which is 2 here.

[0076] The detailed calculation process for the theoretical values ​​and ranges of the four key operating parameters is as follows:

[0077] Vibration intensity ( Calculation of physical saturation coefficient: Millimeters per second. The weight parameters obtained after model training are: load weights. Temperature weighting bias Substitute the values ​​into the formula to calculate the intermediate value of the exponent: Calculate the exponent in the denominator: Final predicted value: Millimeters per second. Calculate the prediction interval: The standard deviation of the residuals of the statistical model on the validation set is 0.15 millimeters per second; calculate the width of fluctuation as three times the standard deviation. The baseline range for determining the vibration intensity threshold is as follows: .

[0078] Current fluctuation ( Calculation of physical saturation coefficient: Ampere. The weight parameters obtained after model training are: load weights (Current magnitude depends primarily on the load), temperature weighting (Ambient temperature has little effect), bias Substitute the values ​​into the formula to calculate the intermediate value of the exponent: Calculate the exponent in the denominator: Final predicted value: Amperes. Calculate the prediction interval: The standard deviation of the residuals of the statistical model on the validation set is 4.0 amperes. Calculate the width of fluctuation of 3 standard deviations. The baseline range for determining the current fluctuation threshold is: .

[0079] Temperature rise range ( Calculation of physical saturation coefficient: Degrees Celsius. The weight parameters obtained after model training are: load weights. Temperature weighting (Ambient temperature has a significant additive effect on the final temperature rise), bias Substitute the values ​​into the formula to calculate the intermediate value of the exponent: Calculate the exponent in the denominator: Final predicted value: Degrees Celsius. Calculate the prediction interval: The standard deviation of the residuals of the statistical model on the validation set is 2.0 degrees Celsius. Calculate the width of fluctuation, which is 3 times the standard deviation. The baseline range for the temperature rise threshold was determined as follows: .

[0080] Fluid pressure ( Calculation of physical saturation coefficient: Megapascals. The weight parameters obtained after model training are: load weights. (Pump outlet pressure is highly correlated with load / speed), temperature weighting bias Substitute the values ​​into the formula to calculate the intermediate value of the exponent: Calculate the exponent in the denominator: Final predicted value: MPa. Calculate the prediction interval: The standard deviation of the residuals of the statistical model on the validation set is 0.5 MPa. Calculate the width of fluctuation of 3 standard deviations. The baseline range for determining the fluid pressure threshold is... .

[0081] S412: Collect maintenance records of the target power plant equipment, combine them with the cumulative operating time of the target power plant equipment, quantify the current overall health level of the target power plant equipment, and establish the current health baseline of the equipment.

[0082] The overall health level of the target power plant equipment is quantified as follows:

[0083] set an initial health degree value of the target power plant equipment, extract a time stamp of the end of the last maintenance from the maintenance record of the target power plant equipment, calculate a time span from the time stamp to the current time as a specified continuous operation duration;

[0084] traverse the maintenance record of the target power plant equipment, count the number of fault maintenance events, and match the corresponding health loss weight value for each fault maintenance event according to the fault severity level recorded in the maintenance record;

[0085] calculate an aging depreciation coefficient of the whole life cycle based on the ratio of the cumulative operation duration of the target power plant equipment to the rated design life of the equipment;

[0086] calculate the performance degradation amount using the specified continuous operation duration and the preset fatigue accumulation rate, determine the damage residual amount using the weighted calculation result of the health loss weight and the number of unplanned fault maintenance events, and calculate the aging loss amount using the aging depreciation coefficient;

[0087] subtract the performance degradation amount, the damage residual amount and the aging loss amount from the initial health degree value in sequence to determine the remaining value as the quantitative result of the current overall health level of the target power plant equipment;

[0088] First, set the initial health degree value to 100, obtain the rated design life of the equipment as 100000 hours (from the equipment factory manual), and the current cumulative operation duration as 12000 hours (from the operation log). Calculate the aging loss amount, adopt the linear depreciation logic, the aging depreciation coefficient calculation formula is , set the aging weight coefficient to 20 (determined based on the equipment life cycle cost analysis), and the aging loss amount is . Calculate the performance degradation amount, extract the continuous operation duration of the last maintenance from the current time as 500 hours, set the fatigue accumulation rate to 0.01 per hour (based on material fatigue experimental data), and the performance degradation amount calculation formula is . Calculate the damage residual amount, consult the maintenance record database, find that there has been 1 "general fault" (the weight defined by historical data is 2.0) and 2 "minor faults" (the weight is 0.5), and the damage residual amount calculation formula is . Finally, quantify the overall health level, subtract each loss from the initial value in sequence, and the calculation formula is , and the value 89.6 is the current health baseline of the equipment.

[0089] S413: Calculate the threshold adjustment ratio required to trigger the warning in the threshold baseline according to the feature similarity score and the current health baseline of the equipment, and perform adjustment on the threshold baseline using the threshold adjustment ratio to obtain the dynamic early warning judgment benchmark;

[0090] First, the threshold adjustment ratio is calculated. The feature similarity score 0.9916 (representing high risk) obtained in step S312 and the health baseline 89.6 (representing sub-health) obtained in step S412 are used. The feature-based adjustment factor calculation rule is set as 1 minus (similarity score multiplied by 0.5), and the calculation result is , which means that the threshold needs to be shrunk by about half due to the high feature being similar to a fault; the health-based adjustment factor calculation rule is set as health baseline divided by 100, and the calculation result is , which means that the threshold needs to be shrunk by about 10% due to the device being slightly aged. The threshold adjustment ratio is calculated by taking a weighted average of the two (the feature weight is set to 0.7 and the health weight is set to 0.3, and the weight allocation is determined by an expert system optimization), and the calculation formula is , and the result is 0.62 after rounding to two decimal places. Finally, the adjustment is performed, and the one-sided floating width of the threshold baseline interval of each key operating parameter determined in step S411 (vibration intensity 0.45, current fluctuation 12.0, temperature rise amplitude 6.0, and fluid pressure 1.5) is compressed using the adjustment ratio. The new one-sided width is calculated as follows: vibration intensity ; current fluctuation ; temperature rise amplitude ; and fluid pressure . The updated dynamic early warning judgment criteria are as follows: 1. vibration intensity: the center value 4.5 remains unchanged, and the new interval is , i.e. ; 2. current fluctuation: the center value 119.7 remains unchanged, and the new interval is , i.e. ; 3. temperature rise amplitude: the center value 40.0 remains unchanged, and the new interval is , i.e. ; and 4. fluid pressure: the center value 15.0 remains unchanged, and the new interval is , i.e. .

[0091] Please refer to Figure 6 , and the steps for obtaining the device fault early warning and diagnosis results are as follows:

[0092] S511: select the predicted deviation values marked by the predicted deviation indicators from the parameter trend prediction data, calculate the ratio of the selected predicted deviation values to the adjusted threshold baseline boundary values in the dynamic early warning judgment criteria, obtain the deviation severity index, and combine the deviation severity index with the feature similarity score to form a decision input vector;

[0093] First, the data is screened, and the labeled data item is extracted from the parameter trend prediction data, with a prediction deviation value of 0.12 (from the calculation in step S213 for the vibration intensity parameter). Then, the deviation severity index is calculated, and the one-sided width of the dynamic early warning judgment criterion for the parameter (vibration intensity) calculated in step S413 is 0.279. The ratio of the calculated deviation value to the reference width is calculated, and the calculation formula is , which means that the current deviation occupies 43% of the allowed alarm margin. The severity state is set to contain two possibilities: if the calculation result is less than or equal to 1.0, it belongs to the non-exceeding range; if the calculation result is greater than 1.0, it belongs to the exceeding range. The current 0.43 belongs to the non-exceeding but risky interval. If the deviation value is 0.3, the calculation result is , which belongs to the exceeding range. The characteristic similarity score is integrated, and the result 0.9916 from step S312 is directly referenced. Finally, the vector is constructed, combining the deviation severity index 0.43 and the characteristic similarity score 0.9916 to form a two-dimensional decision input vector [0.43, 0.9916], which digitally represents the complex state of "although the numerical value has not been severely exceeded (0.43), the fault characteristic form is extremely obvious (0.99)".

[0094] S512: input the preset fault state label as the label input, construct the relationship weight mapping between the decision input vector and the label input, use the fuzzy classification algorithm to determine the early warning state of the decision input vector according to the relationship weight mapping, and output the early warning state determination result;

[0095] First, define the fuzzy reasoning rules, set the input variable "severity" to have three possible categories: low severity (corresponding to the numerical interval [0, 0.5]), medium severity (corresponding to the numerical interval [0.5, 1.0]), and high severity (corresponding to the numerical interval [1.0, positive infinity]); set "similarity" to have two possible categories: low similarity (corresponding to the numerical interval [0, 0.8]) and high similarity (corresponding to the numerical interval [0.8, 1.0]). Establish the rule weight: rule 1 "if the severity is low and the similarity is high, determine early warning", weight 0.8; rule 2 "if the severity is high and the similarity is high, determine serious fault", weight 1.0. Perform multi-value logic reasoning on the decision input vector [0.43, 0.9916]: for the severity 0.43, the membership degree to the "low severity" interval is 1.0 (judged according to the numerical value being completely within the interval), and for the similarity 0.9916, the membership degree to the "high similarity" interval is 1.0. According to rule 1, the intensity of triggering "early warning" is calculated, and the calculation formula is According to rule 2, since the severity level is not "high severity", the trigger strength is 0. Finally, defuzzification is performed, and the centroid method is used to calculate the final discretized decision value, based on the calculation formula. In the formula The risk score is the defuzzified result (representing the system's final quantitative assessment of the danger level of the current state). The first sampled for the risk domain A discrete point value (covering the interval [0, 1], taking...) ), This represents the aggregation membership degree corresponding to this point (reflecting the output strength after the rule is triggered; a larger value indicates a higher confidence level of the inference engine in the risk point. Since "early warning" is triggered and the center of this set is around 0.5, the value is [value missing]. ), This represents the number of sampling points (set to 3 here to cover low, medium, and high segments). Substituting this into the numerator of the numerical calculation... The denominator is calculated as follows: Final calculation The value falls precisely at the center of the early warning set, and the system outputs the judgment result as "early warning triggered".

[0096] S513: When the warning status determination result is to trigger the warning, the deviation severity index and feature similarity score in the decision input vector are numerically sorted, the object with the largest value is determined as the dominant feature, the dominant anomaly source is determined according to the dominant feature, the fault type is determined by combining the fault feature matching result, and the equipment fault warning and diagnosis result is output.

[0097] First, diagnosis is initiated when the judgment result is "triggered early warning". The components in the decision input vector are compared, with the deviation severity index (0.43) and feature similarity score (0.9916) numerically compared. Two possibilities for feature dominance are defined: if the severity index is greater than the similarity score, it is determined to be amplitude exceeding the limit dominance; if the similarity score is greater than the severity index, it is determined to be morphological matching dominance. Because... , determine the dominant feature as the possibility interval of the shape matching, and determine the abnormal source. Since the dominant feature is the similarity score, the internal logic tree of the system determines that the abnormality is not caused by a simple energy surge (such as foreign object impact), but by the pathological evolution of the operation mode, so the dominant abnormal source is determined as "fault mechanism evolution matching". The output result is integrated, the feature matching result "bearing wear" of step S313 is called, the dominant source analysis is combined, and the final diagnosis report is generated: "the device is currently in the early evolution stage of bearing wear, although the vibration amplitude has not yet seriously exceeded the standard (severity 0.43), the multidimensional feature vector has a very high similarity (0.99) with the bearing wear mode, and the device health has decreased. It is recommended to focus on checking the inner ring wear of the non-drive end bearing during the next shutdown, and immediate shutdown is not required but close monitoring is needed."

[0098] The above is only a preferred embodiment of the present application, and does not limit the present application in other forms. Any skilled person in the art can modify or change the above disclosed technical content to equivalent embodiments applied to other fields, but any simple modification, equivalent change and modification made to the above embodiments without departing from the technical solution content of the present application, according to the technical essence of the present application, still belongs to the protection scope of the technical solution of the present application.

Claims

1. A trend-based fault prediction method based on dynamic pattern and threshold coordination, characterized in that, Includes the following steps: S1: Collect key operating parameters and current operating condition parameters of the target power plant equipment within a specified continuous operating cycle, classify and process them, and obtain a sequence of equipment operating status data. S2: Determine the reference data segment in the equipment operating status data sequence, estimate the predicted values ​​of key operating parameters at future time points based on the reference data segment, calculate the prediction deviation between the predicted value and the corresponding current value, and select parameter trend prediction data; S3: Construct multiple feature vectors for the target power plant equipment based on the equipment operation status data sequence, compare them with the stored fault mode feature set, calculate the feature similarity score, and obtain the fault feature matching result; S4: Perform combined regression analysis on the current operating condition parameters in the equipment operating status data sequence, and determine the dynamic early warning judgment benchmark by constructing a backpropagation neural network to fit the benchmark floating curve of each current operating condition parameter under multiple operating conditions. S5: Calculate the severity index of the deviation between the predicted deviation value and the dynamic early warning judgment benchmark, and combine it with the normalized feature similarity score to determine the early warning status, thereby obtaining the equipment fault early warning and diagnosis results.

2. The trend fault prediction method based on dynamic pattern and threshold coordination according to claim 1, characterized in that, The equipment operating status data sequence includes vibration intensity data, current fluctuation data, temperature rise amplitude data, fluid pressure data, load rate, ambient temperature data, cumulative running time, and spindle speed data arranged in chronological order. The parameter trend prediction data includes prediction deviation values ​​and marked prediction offset indicators. The fault feature matching results include feature similarity scores and updated existing feature sets. The dynamic early warning judgment benchmark includes a threshold baseline generated based on a benchmark floating curve and a threshold adjustment ratio. The equipment fault early warning and diagnosis results include early warning status judgment and identified abnormal sources.

3. The trend fault prediction method based on dynamic pattern and threshold coordination according to claim 1, characterized in that, The specific steps for obtaining the device operating status data sequence are as follows: S111: Collect vibration intensity, current fluctuation, temperature rise and fluid pressure data of target power plant equipment within a specified continuous operating cycle to form key operating parameters, construct time series data of corresponding key operating parameters, compare key operating parameters with preset physical limit thresholds to identify abnormal extreme values, remove abnormal extreme values ​​in time series data, and generate key operating parameter time series. S112: Obtain the load rate of the target power plant equipment, the temperature data of the operating environment of the target power plant equipment, the cumulative running time of the target power plant equipment, and the spindle speed data of the target power plant equipment to form the current operating condition parameters. Map the current operating condition parameters to a unified time dimension, perform correlation processing on the current operating condition parameters with inconsistent dimensions, and establish a set of current operating condition parameters. S113: Based on the current operating condition parameter set, classify the operating condition categories of various power plant equipment, and assign each key operating parameter in the key operating parameter time series to the corresponding operating condition category to obtain the equipment operating status data sequence.

4. The trend fault prediction method based on dynamic pattern and threshold coordination according to claim 3, characterized in that, The specific steps for obtaining the parameter trend prediction data are as follows: S211: Extract the latest data of a specified duration from the equipment operation status data sequence to form a real-time monitoring data segment, calculate the numerical difference between the real-time monitoring data segment and the data segment of the same duration in the equipment operation status data sequence within the specified time interval, and determine the data segment with a numerical difference less than a preset matching threshold as the benchmark data segment. S212: Based on the numerical changes of the reference data segment, estimate the predicted value of each key operating parameter at several future time points, calculate the correction compensation amount of the predicted value and perform numerical correction on the predicted value to generate the corrected predicted value of the key operating parameter. S213: Calculate the difference between the current value of the key operating parameter in the equipment operating status data sequence and the corresponding corrected predicted value of the key operating parameter to obtain the prediction deviation value. Mark the parameter values ​​whose prediction deviation values ​​exceed 95% of the numerical distribution range in the equipment operating status data sequence as prediction offset indicators. Integrate the prediction deviation value and the prediction offset indicator to generate parameter trend prediction data.

5. The trend fault prediction method based on dynamic pattern and threshold coordination according to claim 4, characterized in that, The specific steps for obtaining the fault feature matching results are as follows: S311: Calculate the vibration amplitude change rate of the target power plant equipment based on the vibration intensity in the equipment operation status data sequence, calculate the current harmonic offset value of the target power plant equipment based on the current fluctuation in the equipment operation status data sequence, calculate the temperature rise gradient data of the target power plant equipment based on the temperature rise amplitude in the equipment operation status data sequence, and construct the vibration amplitude change rate, current harmonic offset value and temperature rise gradient data into a feature vector to obtain a multi-dimensional operation feature vector; S312: Obtain the stored fault mode feature set, normalize the multidimensional operating feature vector and the stored fault mode feature set, calculate the cosine similarity between the normalized multidimensional operating feature vector and the normalized stored fault mode feature set, obtain the quantitative score of feature matching based on the cosine similarity, and generate a feature similarity score. S313: Based on the feature similarity score, determine whether the current state matches a known fault mode in the stored fault mode feature set, identify new potential fault mode features from the device operating state data sequence and update the stored fault mode feature set to obtain a fault feature matching result.

6. The trend fault prediction method based on dynamic pattern and threshold coordination according to claim 5, characterized in that, The stored fault mode feature set includes: a standard fault sample vector library after historical operation and maintenance data cleaning and calibration, and corresponding fault category definition labels; Each sample vector in the standard fault sample vector library is composed of historical vibration amplitude change rate, historical current harmonic offset value and historical temperature rise gradient data collected under confirmed fault conditions, arranged in the same dimensional order when constructing the multidimensional operating feature vector. The fault category definition labels include vibration anomaly labels that characterize the degree of mechanical wear of rotating components, current anomaly labels that characterize the electrical insulation state of the stator and rotor, and temperature rise anomaly labels that characterize the cooling efficiency of the heat exchange process.

7. The trend fault prediction method based on dynamic pattern and threshold coordination according to claim 5, characterized in that, The specific steps for obtaining the dynamic early warning judgment criteria are as follows: S411: Perform multivariate regression analysis on the current operating condition parameter set in the equipment operating status data sequence, and fit the benchmark floating curve of each current operating condition parameter under multiple operating conditions by constructing a backpropagation neural network. Input the current operating condition parameter set into the benchmark floating curve, calculate the theoretical prediction interval of key operating parameters, and use the theoretical prediction interval of key operating parameters as the threshold baseline of the target power plant equipment under the current operating state. S412: Collect maintenance records of the target power plant equipment, combine them with the cumulative operating time of the target power plant equipment, quantify the current overall health level of the target power plant equipment, and establish the current health baseline of the equipment. S413: Calculate the threshold adjustment ratio required to trigger an early warning in the threshold baseline based on the feature similarity score and the current health baseline of the device, and adjust the threshold baseline using the threshold adjustment ratio to obtain a dynamic early warning judgment benchmark.

8. The trend fault prediction method based on dynamic pattern and threshold coordination according to claim 7, characterized in that, The overall health level of the target power plant equipment currently includes: Set the initial health value of the target power plant equipment, extract the timestamp of the most recent maintenance completion from the maintenance records of the target power plant equipment, and calculate the time span from the timestamp to the current time as the specified continuous operation duration; Iterate through the maintenance records of the target power plant equipment, count the number of fault maintenance events, and match the corresponding health loss weight value for each fault maintenance event based on the fault severity level recorded in the maintenance record. The aging depreciation factor for the entire life cycle is calculated based on the ratio of the cumulative operating time of the target power plant equipment to the rated design life of the equipment. The performance degradation is calculated using a specified continuous working duration and a preset fatigue accumulation rate. The residual damage is determined by a weighted calculation of the number of unplanned failure maintenance events and the health loss weight. The aging loss is calculated using the aging depreciation factor. The initial health value is subtracted sequentially from the performance degradation, residual damage, and aging loss. The remaining value is then used as the quantitative result of the current overall health level of the target power plant equipment.

9. The trend fault prediction method based on dynamic pattern and threshold coordination according to claim 7, characterized in that, The formula used to calculate the theoretical prediction range of key operating parameters is as follows: ； in, For the first Theoretical predicted values ​​of key operating parameters, For the first The physical saturation coefficient of each parameter, The base of the natural logarithm, For the first The input condition features for the first The weights of each output parameter, For the first Each input working condition feature value In the multivariate regression model, the first Activation bias of each output neuron For input dimensions.

10. The trend fault prediction method based on dynamic pattern and threshold coordination according to claim 7, characterized in that, The specific steps for obtaining the equipment fault early warning and diagnosis results are as follows: S511: Filter the prediction deviation values ​​marked by the prediction offset index from the parameter trend prediction data, calculate the ratio of the filtered prediction deviation values ​​to the adjusted threshold baseline boundary value in the dynamic early warning judgment benchmark, obtain the deviation severity index, and normalize and combine the deviation severity index with the feature similarity score to construct a decision input vector. S512: Using the preset fault status label as the label input, construct the relationship weight mapping between the decision input vector and the label input, use the fuzzy classification algorithm to determine the warning status of the decision input vector according to the relationship weight mapping, and output the warning status determination result. S513: When the warning status determination result is a triggered warning, the deviation severity index and the feature similarity score in the decision input vector are numerically sorted, the object with the largest value is determined as the dominant feature, the dominant anomaly source is determined according to the dominant feature, the fault type is determined by combining the fault feature matching result, and the equipment fault warning and diagnosis result is output.

Images