Multi-source data fusion monitoring method for the operating status of secondary control equipment in hydropower stations

Abstract

This invention relates to the field of hydropower station equipment monitoring, and discloses a multi-source data fusion monitoring method for the operating status of secondary industrial control equipment in hydropower stations. The method includes: collecting multi-channel data on vibration, current, and temperature to construct an operating sample; reconstructing a time-series operating data matrix, constructing a cooling circuit temperature difference degradation index and vibration-current coupling characteristics, concatenating these with channel-based statistics to form a physical feature vector, and performing channel-level standardization on the time-series data; extracting multi-scale time-series features using parallel convolutional branches, and using the physical feature vector to generate a gating vector to recalibrate the multi-scale feature map channels, outputting the fault probability after classification; constructing a physical matching vector using the physical feature vector to perform consistency correction on the fault probability, and combining the sliding window trend of the temperature difference degradation index with the judgment confidence to generate an early warning index, based on which an alarm decision is output. This invention integrates physical priors and deep time-series features, improving the accuracy and anti-interference capability of early fault identification.

Description

Technical Field

[0001] This invention relates to the field of hydropower station equipment monitoring, specifically a method for multi-source data fusion monitoring of the operating status of secondary industrial control equipment in hydropower stations. Background Technology

[0002] Secondary control equipment in hydropower stations integrates complex mechanical transmissions, electrical circuits, and cooling systems. It is a critical link in ensuring the safe operation of generator units, and its operating status directly affects the continuity of power production and equipment safety. This type of equipment typically contains complex mechanical, electrical, and cooling systems, and its failures often undergo a gradual process from early minor deterioration to eventual failure. Therefore, online condition monitoring and early fault identification are of great significance for achieving condition-based maintenance and avoiding unplanned downtime.

[0003] However, existing technical solutions still have significant shortcomings in three aspects: fault feature extraction, multi-source information fusion, and alarm decision-making mechanisms. Specifically, these shortcomings are as follows: (1) Existing methods often directly normalize or differentially process multi-channel time series data, lacking differentiated feature construction for specific fault mechanisms of secondary industrial control equipment in hydropower stations, and cannot effectively capture early weak deterioration signs such as continuous expansion of cooling circuit temperature difference and vibration impact under fluctuating current conditions.

[0004] (2) Traditional multi-scale convolutional networks assign the same weight to different samples, and the handmade features are simply spliced ​​at the end of the classification layer. They lack a mechanism to actively adjust the importance of temporal feature channels before feature aggregation by utilizing physical priors.

[0005] (3) Most methods rely on the model output probability of a single sample to trigger an alarm, without introducing physical quantities directly corresponding to the fault mechanism for consistency verification, and also lack modeling of the time continuity of the deterioration trend, which can easily lead to misjudging instantaneous operating condition fluctuations as faults.

[0006] In summary, existing technologies urgently need to achieve breakthroughs in three areas: mechanism-driven feature engineering, physical prior-guided attention fusion mechanisms, and robust alarm strategies based on physical quantity consistency verification and degradation trend modeling, in order to truly meet the high reliability and timeliness requirements of online monitoring for secondary industrial control equipment in hydropower stations. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a multi-source data fusion monitoring method for the operating status of secondary industrial control equipment in hydropower stations, thereby resolving the technical problems mentioned in the background section.

[0008] To achieve the above objectives, this invention employs the following technical solution: a multi-source data fusion monitoring method for the operating status of secondary control equipment in hydropower stations, comprising the following steps: S1. Collect multi-channel raw operating data during the continuous operation of secondary industrial control equipment of hydropower station. The multi-channel raw operating data includes the effective value of vibration acceleration, the effective value of A-phase current, oil temperature and cooling water temperature. Multiple consecutive sampling time steps constitute an operating sample, and a real state category label is configured for each operating sample. S2. Reconstruct the original operating data into a time-series operating data matrix, construct the cooling circuit temperature difference degradation index and vibration-current coupling characteristics, and concatenate the cooling circuit temperature difference degradation index, vibration-current coupling characteristics and the basic statistics of each operating data channel into a physical feature vector. Then, use the channel statistics of the normal class training samples to perform channel-level standardization on the time-series operating data matrix to obtain the standardized time-series operating data matrix. S3. Construct a multi-scale temporal convolutional network, use multiple parallel one-dimensional dilated convolutional branches to extract temporal features at different time scales from the standardized temporal running data matrix to obtain a multi-scale temporal feature map, and use physical feature vectors to generate gating vectors, perform channel-by-channel recalibration on the multi-scale temporal feature map, and then output the fault probability vector through the classification layer. S4. Construct a physical matching vector using physical feature vectors, perform category-by-category consistency correction on the fault probability vector to obtain a cooperative probability vector, and determine the final state category and final judgment confidence level based on the cooperative probability vector; at the same time, perform sliding window trend analysis on the cooling circuit temperature difference deterioration index of multiple consecutive operating samples, integrate the final judgment confidence level with the positive change trend of the cooling circuit temperature difference deterioration index to obtain the early warning index, and output alarm decision based on the early warning index.

[0009] Furthermore, the construction methods for the cooling circuit temperature difference degradation index in S2 include: Extract the oil temperature sequence and coolant temperature sequence from the time-series running data matrix, calculate the difference between the oil temperature and coolant temperature at the same time step, and obtain the temperature difference sequence; Perform least-squares linear fitting on the temperature difference sequence to obtain the linear fitting slope and linear fitting intercept, and calculate the temperature difference fitting error; The positive portion of the linear fitting slope is retained, and the positive change rate is mass-weighted using the temperature difference fitting error to obtain the cooling circuit temperature difference degradation index.

[0010] Furthermore, the methods for constructing the vibration-current coupling characteristics in S2 include: Extract the effective value sequence of vibration acceleration and the effective value sequence of phase A current from the time-series operation data matrix, and calculate the vibration mean, vibration standard deviation, current mean and current standard deviation respectively; Calculate the current standard deviation corresponding to each normal class training sample, and perform an arithmetic mean on the current standard deviations of all normal class training samples to obtain a reference value for the current standard deviation. Based on the deviation between the current standard deviation of the current sample and the reference value of the current standard deviation, an adaptive threshold coefficient is generated using the hyperbolic tangent function; The vibration impact detection threshold is determined based on the vibration mean, vibration standard deviation and adaptive threshold coefficient. The effective value of vibration acceleration is compared with the vibration impact detection threshold at each time step, and the number of vibration impacts is counted. Calculate the vibration kurtosis and concatenate the vibration impact number, vibration kurtosis and current standard deviation into a vibration-current coupling feature vector.

[0011] Furthermore, the physical feature vector includes: cooling circuit temperature difference degradation index, vibration impact number, vibration kurtosis, current standard deviation, mean oil temperature, mean coolant temperature, mean vibration, and mean current, and the features in the physical feature vector are arranged in a fixed order.

[0012] Furthermore, S3 constructs a multi-scale temporal convolutional network, utilizing multiple parallel one-dimensional dilated convolutional branches to extract temporal features at different time scales from the standardized temporal running data matrix, resulting in a multi-scale temporal feature map, specifically including: The standardized time-series running data matrix is ​​transposed to obtain the convolution input matrix; The convolution input matrix is ​​simultaneously input into three parallel one-dimensional dilated convolution branches; The time dimension padding is set according to the kernel length and dilation rate of each one-dimensional dilated convolution branch to keep the output time length of the three one-dimensional dilated convolution branches consistent. Batch normalization and ReLU activation are performed sequentially on the convolution output of each one-dimensional dilated convolution branch to obtain the corresponding one-dimensional dilated convolution branch feature map.

[0013] Furthermore, S3 utilizes physical feature vectors to generate gating vectors and performs channel-by-channel recalibration on multi-scale temporal feature maps, specifically including: The physical feature vector is input into the gated fully connected layer to obtain the gated intermediate vector; Perform a sigmoid mapping on each element of the gated intermediate vector to obtain the gated vector. The dimension of the gated vector is the same as the number of channels in the multi-scale temporal feature map. The gated vector is copied along the time dimension and multiplied channel-by-channel with the multi-scale temporal feature map to obtain the gated weighted feature map.

[0014] Furthermore, the methods for constructing the physical matching vector in S4 include: Physical quantities that are physically related to each fault category are selected from the physical feature vector. For mechanical faults, the number of vibration impacts and vibration kurtosis are selected; for cooling system faults, the cooling circuit temperature difference deterioration index is selected; and for electrical faults, the current standard deviation is selected. The normal baseline mean and normal baseline standard deviation of each physical quantity were statistically analyzed using normal training samples. Calculate the standardized deviation of each physical quantity of the current running sample relative to the corresponding normal baseline, and use the biased Sigmoid function to map the standardized deviation to physical matching weights in the range of 0 to 1 to form a physical matching vector.

[0015] Furthermore, in S4, a class-by-class consistency correction is performed on the fault probability vector to obtain the cooperative probability vector, specifically including: The fault probability vector is weighted element-wise using the physical matching vector; The cooperative probability vector is obtained by dividing the weighted probability of each state category by the sum of the weighted probabilities of all state categories.

[0016] Furthermore, the sliding window trend analysis in S4 includes: An exponentially weighted moving average is applied to the cooling circuit temperature difference degradation index sequence to obtain a smoothed cooling circuit temperature difference degradation index. Within the deterioration trend sliding window, weighted linear regression is performed using exponential decay weights to calculate the smoothed deterioration rate of change. The early warning index is obtained by dividing the smoothed degradation rate by the standard deviation of the cooling circuit temperature difference degradation index of the normal training samples, and then multiplying it by the final judgment confidence level after index mapping.

[0017] Furthermore, S5 is also included: After completing the training of the multi-scale temporal convolutional network and determining various fixed parameters, the trainable parameters of the multi-scale temporal convolutional network and the fixed parameters determined by the normal class training samples are deployed to the status monitoring system of the secondary industrial control equipment of the hydropower station; during online operation, multi-channel operation data is continuously collected to form the sample to be detected, and S2, S3 and S4 are executed in sequence. When the warning index is not less than the preset warning threshold, the fault alarm signal and the final status category are output; otherwise, only the status monitoring log is recorded.

[0018] To achieve the above objectives, the present invention employs the following technical solution: Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a cooling circuit temperature difference degradation index and current adaptive vibration and shock detection feature with clear physical meaning. It introduces the linear fitting quality weighting of temperature difference and the dynamic threshold of current conditions into feature engineering, which solves the problem that a single statistical quantity is difficult to distinguish between continuous degradation and random fluctuations.

[0019] 2. This invention constructs a physical feature-gated multi-scale temporal convolutional network, uses manually generated physical features to generate channel-level gating weights, guides the network to adaptively adjust the importance of convolutional feature channels at different time scales, and realizes a deep integration of physical knowledge and deep learning feature extraction.

[0020] 3. This invention adopts a warning decision mechanism that combines physical consistency correction and degradation trend sharpness. It superimposes physical matching constraints and cooling loop trend index mapping on the model probability output, nonlinearly fuses single sample confidence with trend information of continuous time windows, and reduces false alarms caused by occasional interference. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram illustrating the construction process of vibration-current coupling characteristics; Figure 3 This is a schematic diagram of the physical feature vector generation and time-series running data standardization process; Figure 4 This is a schematic diagram of the multi-scale dilated convolutional temporal coding process; Figure 5 This is a schematic diagram of the channel gating and fault classification process; Figure 6 This is a schematic diagram of the gating weight vector for a normal state sample; Figure 7 This is a schematic diagram of the gating weight vector for mechanical faults; Figure 8 This is a schematic diagram of the gating weight vector for cooling system faults. Detailed Implementation

[0022] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0023] like Figure 1 As shown, this invention proposes a multi-source data fusion monitoring method for the operating status of secondary control equipment in hydropower stations, the main contents of which are as follows: S1. Data Acquisition and Sample Construction for Secondary Control Equipment in Hydropower Stations

[0024] This invention collects the effective values ​​of vibration acceleration, A-phase current, oil temperature, and cooling water temperature during the continuous operation of secondary industrial control equipment in hydropower stations. The effective value of vibration acceleration is used to characterize the mechanical vibration intensity of the equipment, the effective value of A-phase current is used to characterize the electrical load and load fluctuation of the equipment, the oil temperature is used to characterize the thermal state of the internal lubricating or insulating medium of the equipment, and the cooling water temperature is used to characterize the temperature state of the cooling medium.

[0025] In one implementation, the four operational data channels are sampled synchronously using a unified clock, with a sampling interval of up to 1 second. Each operational sample consists of 60 consecutive sampling time steps, including 60 RMS values ​​of vibration acceleration, 60 RMS values ​​of phase A current, 60 oil temperature values, and 60 coolant temperature values. Following the sampling time sequence, the RMS values ​​of vibration acceleration, phase A current, oil temperature, and coolant temperature for each time step are arranged sequentially to form an original operational data vector with a dimension of 240.

[0026] Specifically, the first The original running data vector is denoted as ,in, Indicates the sample index. It refers to the first The sampled values ​​of 60 time steps and 4 running data channels in a running sample are arranged in a fixed order to form a vector with a dimension of 240.

[0027] Based on equipment maintenance records, fault test records, or manual confirmation results, assign a true status category label to each raw operational data vector. Specifically, the true status category label is denoted as... , The value ranges from 1 to 4, corresponding to four predetermined equipment operating status categories, among which, Corresponding to the normal state, Corresponding to mechanical failures, Corresponding to cooling system malfunction, Corresponding to electrical faults. Training data, validation data, and data to be tested use the same sampling interval, time window length, channel arrangement order, and category numbering rules.

[0028] In one implementation, the labels are manually labeled. Specifically, historical work order records can be exported from the hydropower station's equipment maintenance management system. Based on the fault conclusions and corresponding time ranges in the work orders, the original operational data vector is generated. Label the actual state category For normal operating periods, the data is marked based on the fault-free operation records.

[0029] S2. Multi-channel degradation feature mining and data preprocessing The four-channel operating data collected by the secondary control equipment of the hydropower station includes the effective value of vibration acceleration, the effective value of phase A current, oil temperature, and cooling water temperature. When an early failure occurs in the cooling system, the temperature difference between the oil and cooling water usually shows a continuously increasing trend; when an early failure occurs in mechanical components, intermittent impacts usually appear in the effective value of vibration acceleration, and the fluctuation of the effective value of phase A current will change the background fluctuation level of the vibration channel. If only normalization or differential processing is performed on the four-channel operating data, it is difficult to simultaneously extract the physical correlation between the temperature difference change trend, the transient impact of vibration, and the fluctuation of current conditions.

[0030] This invention reconstructs each original operational data vector into a time-series operational data matrix, constructing a cooling circuit temperature difference degradation index and vibration-current coupling features. The cooling circuit temperature difference degradation index, vibration-current coupling features, and the basic statistics of the four operational data channels are then concatenated to form a physical feature vector. Simultaneously, channel-level normalization is performed on the time-series operational data matrix using the channel statistics of normal training samples, resulting in a normalized time-series operational data matrix. The physical feature vector guides subsequent multi-scale temporal convolutional networks, and the normalized time-series operational data matrix provides dynamic change information for the four operational data channels over a continuous 60 seconds. The specific steps are as follows: S201. Construction of Cooling Loop Temperature Difference Deterioration Index Based on Linear Fitting and Quality Weighting Cooling system deterioration is usually manifested as a gradual increase in the temperature difference between oil temperature and coolant temperature over a continuous time range. Affected by sensor noise and equipment load fluctuations, the temperature difference at a single time step may fluctuate randomly. It is difficult to distinguish between a continuous trend of increasing temperature difference and occasional temperature fluctuations by only calculating the average temperature difference.

[0031] This invention performs linear fitting on a continuous 60-second temperature difference sequence and constructs a cooling circuit temperature difference degradation index by combining the linear fitting slope and the temperature difference fitting error. This allows operating samples with continuously expanding and stable temperature differences to obtain a larger cooling circuit temperature difference degradation index. The specific steps are as follows: 1> According to the original running data vector Based on the sampling time order and channel arrangement order, the four running data channels corresponding to each time step are extracted into one row, resulting in the first row. The time-series running data matrix corresponding to each running sample Time-series running data matrix The size is The first column corresponds to the effective value of vibration acceleration, the second column corresponds to the effective value of phase A current, the third column corresponds to oil temperature, and the fourth column corresponds to cooling water temperature; The row corresponds to the first Each sampling time step This represents the time step index, with a value range of 1 to 60.

[0032] In practical implementation, when the original running data vector When arranged in the order of "4 channels of the first time step, 4 channels of the second time step, up to 4 channels of the 60th time step", the original running data vector The data is reconstructed from every four adjacent values ​​into a time-series running data matrix. Line 1 in the table. This reconstruction operation only changes the data arrangement; it does not change any sampled value or its corresponding sampling time.

[0033] It should be noted that the training data, validation data, and data to be detected are all reconstructed using the same channel arrangement order to ensure that the meaning of the data remains consistent in subsequent channel extraction, physical feature calculation, and model inference processes.

[0034] 2> From the time-series running data matrix The oil temperature sequence is extracted from the third column, and the cooling water temperature sequence is extracted from the fourth column. Among them, the... The running sample is in the first The oil temperature at each time step is recorded as follows: , Used to characterize the The internal thermal state of the equipment at each time step; the cooling water temperature at the corresponding time step is recorded as... , Used to characterize the The temperature state of the cooling medium at each time step.

[0035] Furthermore, by subtracting the cooling water temperature from the oil temperature at the same time step, a temperature difference sequence is obtained. ,Right now ;in, It refers to the first The running sample is in the first The difference between oil temperature and coolant temperature at each time step. Temperature difference sequence. It is used to reduce the absolute temperature deviation caused by the simultaneous rise or fall of oil temperature and cooling water temperature, and to highlight the trend of oil temperature gradually deviating from cooling water temperature when the heat exchange capacity of the cooling circuit decreases.

[0036] For example, when increased equipment load causes both oil and cooling water temperatures to rise by 2°C, the absolute values ​​of both oil and cooling water temperatures change, but the temperature difference between them may remain essentially unchanged. When the heat exchange capacity of the cooling circuit decreases, the rise in oil temperature is usually greater than the rise in cooling water temperature, and the temperature difference sequence... It will gradually increase over time.

[0037] 3> Indexed by time step As the independent variable, the temperature difference sequence Using the dependent variable, a least-squares linear fit was performed on 60 temperature difference data points to obtain the... The slope of the linear fit corresponding to each running sample and linear fitting intercept .

[0038] In the actual implementation, the average value of the 60 time step indices is first calculated. and the average of 60 temperature difference data ,in, This represents the arithmetic mean of time step indices 1 to 60. Indicates the first The average temperature difference of the sampled data within the current time window is calculated; then, the slope of the linear fit is calculated based on the co-variation relationship between the time step index and the temperature difference data. And based on the slope of the linear fit Calculate the linear fit intercept , represented as: ; ; in, It refers to the slope of the linear fit of the temperature difference sequence in the i-th running sample. It is the average change of temperature difference within a unit time step obtained by least squares linear fitting, and is used to characterize the direction and rate of temperature difference change. This refers to the linear fitting intercept of the temperature difference sequence in the i-th running sample, used to determine the overall position of the fitted line within the temperature difference range. Based on the linear fitting slope... and linear fitting intercept The obtained fitted temperature difference is .

[0039] In practical implementation, when the sampling interval is 1 second, the linear fitting slope is... This corresponds to the average change in temperature per second. When the value is positive, it indicates that the temperature difference between the oil temperature and the coolant temperature is generally increasing; when... A value of 0 or negative indicates that no deterioration trend of continuously widening temperature difference has formed within the current time window.

[0040] 4> Based on linear fitting slope and linear fitting intercept The residuals between the actual temperature difference and the fitted temperature difference are calculated step by step over time, and the average of the squares of the 60 residuals is obtained to obtain the nth residual. Temperature difference fitting error of the sample This characterizes the average deviation of the temperature difference sequence from the fitted straight line, and is used to measure the stability of the current temperature difference change trend. The calculation method is expressed as follows: .

[0041] It should be noted that, The smaller the value, the closer the temperature difference change is to a continuous and stable linear change; The larger the value, the stronger the random fluctuations, short-term jumps, or operating condition disturbances contained in the temperature difference sequence.

[0042] Furthermore, the slope of the linear fit is preserved. The positive part, and the fitting error using temperature difference. Mass-weighted calculation of the positive rate of change yields the cooling circuit temperature difference degradation index. ,Right now .

[0043] in, It refers to the first The cooling circuit temperature difference degradation index of the sample is used to quantify the degree of continuous expansion of the temperature difference between oil temperature and cooling water temperature. Indicates the slope of the linear fit Taking a larger value between 0 and 0, the cooling circuit temperature difference deterioration index is reduced. Only responds to processes where the temperature difference increases; The degradation index is used to calculate the stability constant, which is used to prevent temperature difference fitting errors when the temperature difference sequence completely coincides with the fitted line. A value of 0 is acceptable. .

[0044] It should be noted that when the temperature difference sequence shows a continuous upward trend and the temperature difference data at each time step are close to the fitted straight line, the slope of the linear fitting is... The positive portion is relatively large, resulting in a large temperature difference fitting error. Smaller, cooling circuit temperature difference degradation index The corresponding increase occurs when the temperature difference sequence only exhibits short-duration jumps at a single time step or a few time steps, even if the local temperature difference is large, the temperature difference fitting error will increase accordingly. It will also increase, thereby suppressing the temperature difference degradation index of the cooling circuit. This reduces misjudgments caused by instantaneous temperature disturbances.

[0045] It should also be noted that the cooling circuit temperature difference degradation index It is calculated entirely from the oil temperature and coolant temperature within a single running sample over a continuous 60 seconds, without including parameters that need to be determined through model training. By utilizing the rate of temperature difference change and the stability of the temperature difference trend, it can distinguish between the continuous expansion of temperature difference caused by the slow deterioration of the cooling system and random temperature fluctuations.

[0046] S202. Construction of Statistical Features of Vibration and Shock Based on Current Adaptive Threshold When mechanical components experience early failures, short-duration impact peaks typically appear in the effective value sequence of vibration acceleration. Different equipment loads and electrical conditions can cause fluctuations in the effective value of phase A current, which in turn changes the background fluctuation level of the vibration channel. When a vibration shock detection threshold is set using a fixed multiple of the vibration standard deviation, normal operating disturbances can easily be identified as vibration shocks under conditions with large current fluctuations.

[0047] like Figure 2 As shown, this invention dynamically adjusts the vibration and shock detection threshold based on the fluctuation degree of the effective value sequence of phase A current, and constructs vibration-current coupling characteristics by combining vibration kurtosis. The specific steps are as follows: 1> From the time-series running data matrix The first column is used to extract the effective value sequence of vibration acceleration, and the second column is used to extract the effective value sequence of phase A current.

[0048] Among them, the The running sample is in the first The effective value of vibration acceleration at each time step is denoted as , Used to characterize the The mechanical vibration intensity at each time step; the effective value of the A-phase current at the corresponding time step is denoted as... , Used to characterize the Electrical load level at each time step.

[0049] Furthermore, the effective values ​​of vibration acceleration at 60 time steps were statistically analyzed to obtain the... Vibration mean of the running sample and vibration standard deviation The effective values ​​of phase A current at 60 time steps were statistically analyzed to obtain the... Average current of the sampled lines and current standard deviation Among them, the vibration mean The baseline level used to characterize the effective value of vibration acceleration within the current time window, the vibration standard deviation. Used to characterize the effective value of vibration acceleration relative to the vibration mean. The degree of fluctuation; average current Current standard deviation is used to characterize the average electrical load level within the current time window. It is used to characterize the fluctuation of the effective value of phase A current within the current time window.

[0050] 2> Predetermine the reference value of the current standard deviation under normal operating conditions based on normal training samples. Reference value for standard deviation of current It is used to characterize the typical fluctuation level of the effective value of phase A current within a 60-second time window under normal operating conditions.

[0051] In the specific implementation, the standard deviation of the current corresponding to each normal class training sample is calculated separately. Then, the arithmetic mean of the current standard deviations of all normal training samples is calculated to obtain the reference value of the current standard deviation. .

[0052] It should be noted that the current standard deviation reference value It is determined only based on normal training samples, remains unchanged after the model training is completed, and is directly called in the verification stage and the inference stage of the sample to be tested, so that the current fluctuation of different running samples can be compared with the same normal operating condition benchmark.

[0053] It should also be noted that the current standard deviation reference value It should be a positive number. In actual equipment operation, the effective value of phase A current usually fluctuates naturally due to load and measurement noise. Therefore, the reference value of the current standard deviation obtained from the statistical analysis of normal training samples... Not zero.

[0054] 3> Set the current standard deviation of the current sample. Reference value of current standard deviation By comparing the results and using the hyperbolic tangent function to limit the influence range of current fluctuation deviation on the threshold adjustment, the first... Adaptive threshold coefficients for each running sample Used to control the vibration and shock detection threshold relative to the vibration standard deviation The offset factor is calculated as follows: ; in, This represents the mapping amplitude adjustment factor, used to limit the threshold adjustment amplitude caused by current fluctuations; it can be set to 0.5. This represents the hyperbolic tangent function, used to compress the relative deviation of the current standard deviation to a minimum. scope.

[0055] It should be noted that when When the value is 0.5, the adaptive threshold coefficient The value range is 2.5 to 3.5. When the current standard deviation... Equal to the reference value of current standard deviation When the input to the hyperbolic tangent function is 0, the adaptive threshold coefficient... The value is 3. When the current fluctuation of the currently operating sample is higher than the normal operating condition baseline, the adaptive threshold coefficient is [value missing]. A value greater than 3 increases the vibration and shock detection threshold, reducing false vibrations and shocks caused by current fluctuations. When the current fluctuation of the current sample is lower than the normal operating condition baseline, the adaptive threshold coefficient... A value less than 3 ensures that the detection process remains sensitive to weak real vibration impacts.

[0056] It should be noted that the hyperbolic tangent function is used to limit the adaptive threshold coefficient. The range of variation can avoid the standard deviation of the current. The vibration and shock detection threshold changes excessively due to short-term strong fluctuations, while retaining the adjustment effect of the effective value of the A-phase current on the vibration background fluctuation level.

[0057] 4> Based on vibration mean Determine the vibration baseline level of the current operating sample based on the vibration standard deviation. Determine the natural fluctuation range of the vibration, and then use the adaptive threshold coefficient. Adjusting the magnification factor of the natural vibration range, we obtain the first... Vibration and shock detection threshold for running samples ,Right now .

[0058] in, It refers to the first The vibration and shock detection threshold corresponding to each operating sample under the current current condition. With vibration mean As a baseline level, the standard deviation of vibration As a background fluctuation scale, and through an adaptive threshold coefficient The influence of the effective value of phase A current on the operating conditions.

[0059] Furthermore, the effective values ​​of vibration acceleration are compared step by step over time. Vibration and shock detection threshold When the effective value of vibration acceleration Greater than the vibration and shock detection threshold When the vibration acceleration reaches its effective value, the corresponding time step is counted as one vibration impact; Not greater than the vibration and shock detection threshold Vibration shocks are not considered at this time. After traversing 60 time steps, the number of vibration shocks is obtained. , represented as: ; in, Indicates the first The number of vibration impacts of each running sample within the current time window is used to characterize the number of transient vibration events that exceed the background fluctuation range corresponding to the current current condition; This indicates a conditional indicator function. It takes the value 1 when the condition in parentheses is true, and 0 otherwise.

[0060] It should be noted that when the vibration standard deviation When the value is 0, it indicates that the effective value of vibration acceleration remains unchanged within the current time window. At this time, there are no impact events exceeding the vibration background fluctuation range. The number of vibration impacts is then... Set to 0.

[0061] 5> Based on the effective value of vibration acceleration relative to the mean vibration value Calculation of vibration kurtosis using the fourth-order normalized central moment The degree of peak concentration in the distribution of effective vibration acceleration values ​​is obtained, and is expressed as: ; in, It refers to the first The kurtosis of a sample is used to characterize the concentration of peak data in the effective value sequence of vibration acceleration relative to the overall fluctuation. The larger the value, the more pronounced the vibration peaks appear within a small number of time steps; vibration kurtosis A smaller value indicates a relatively flat distribution of the effective values ​​of vibration acceleration. The same definition of vibration kurtosis is used in the training, validation, and inference phases for the samples to be tested. When the vibration standard deviation... When it is 0, the vibration kurtosis Set to 0.

[0062] Furthermore, the number of vibration impacts Vibration kurtosis and current standard deviation By splicing them together in a fixed order, the vibration-current coupling characteristic vector is obtained. ,Right now .

[0063] in, It refers to the first The vibration-current coupling feature vector of the running sample, with a dimension of 3; superscript Represents vector transpose. Vibration-current coupling eigenvector. It also includes the number of vibration impacts, the degree of vibration peaks, and the level of current fluctuations, so that subsequent models can combine the operating conditions provided by the electrical channels when judging the mechanical state.

[0064] It should be noted that the number of vibration and impact cycles... The kurtosis is used to reflect the number of peaks exceeding the adaptive background threshold. This is used to reflect the peaks in the vibration data distribution. Combining the two can reduce the problem of being overly sensitive to individual abnormal sampling points when relying solely on the number of vibration impacts; and, the current standard deviation It participates in both vibration and shock detection threshold adjustment and serves as a vibration-current coupling feature vector. The components are preserved, enabling subsequent models to further learn the relationship between current condition fluctuations and vibration anomalies.

[0065] S203, Physical Feature Vector Generation and Time-Series Operation Data Standardization Cooling circuit temperature difference degradation index and vibration-current coupling eigenvectors They respectively reflect the deterioration trend of the cooling system and the characteristics of mechanical shock, but cannot fully characterize the basic operating level of the four operating data channels within the current time window.

[0066] This invention measures the temperature difference degradation index of the cooling circuit. Vibration-current coupling characteristic vector The data is concatenated with the basic statistics of the four running data channels to obtain a sample-level physical feature vector; simultaneously, the channel statistics of the normal class training samples are used to modify the time-series running data matrix. Implement standardization, such as Figure 3 As shown, the specific steps are as follows:

[0067] 1> For time-series running data matrix The 60 oil temperature values ​​in the third column are averaged to obtain the average oil temperature. ; Perform an arithmetic mean on the 60 cooling water temperature values ​​in column 4 to obtain the average cooling water temperature. Meanwhile, the vibration mean value obtained from S202 is used. and average current This forms the basic level statistics for four operational data channels.

[0068] Among them, the average oil temperature Used to characterize the overall level of oil temperature within the current time window; average coolant temperature. Used to characterize the overall temperature level of the cooling medium within the current time window; vibration mean. The baseline level used to characterize equipment vibration within the current time window; mean current. Used to characterize the average electrical load level of equipment within the current time window.

[0069] 2> The temperature difference degradation index of the cooling circuit Number of vibration and impact cycles Vibration kurtosis Current standard deviation Average oil temperature Average cooling water temperature Vibration mean and average current By concatenating them in a fixed order, we obtain the physical feature vector. ,Right now .

[0070] in, It refers to the first The physical feature vector of each running sample has a dimension of 8. It also characterizes the trend of cooling circuit deterioration, vibration transient impact, current condition fluctuations, and the basic operating level of four operating data channels.

[0071] It should be noted that physical feature vectors The order of the eight physical features remains consistent throughout the model training, validation, and inference phases to avoid changes in the input meaning of the gated fully connected layer due to changes in feature positions.

[0072] 3> Based on the normal class training samples, calculate the mean and standard deviation of the training set channels for each running data channel. Specifically, for the first... Each running data channel aggregates the data from 60 time steps of all normal class training samples in that channel, and calculates the channel mean of the training set. and the standard deviation of the training set channels .

[0073] in, This represents the index of the running data channel, with values ​​ranging from 1 to 4, corresponding to the effective value of vibration acceleration, the effective value of phase A current, oil temperature, and cooling water temperature, respectively; the average value of the training set channels. Used to characterize the first under normal operating conditions The baseline level of each running data channel; the standard deviation of the training set channels. Used to characterize the first under normal operating conditions The natural fluctuation range of each running data channel.

[0074] It should be noted that the training set channel mean and the standard deviation of the training set channels The parameters are determined solely based on normal training samples and remain fixed during the model validation and inference phases for samples to be tested. They are not recalculated using validation or samples to be tested, thus avoiding the use of validation or samples to be tested in the standardized parameter estimation.

[0075] 4> Targeting the time-series running data matrix For each running data channel, the mean of the corresponding training set channels is used. and the standard deviation of the training set channels Perform Z-Score standardization. Specifically, the Z-Score will be... Subtract the mean of the training set channels from each sample value in the running data channels. Then divide by the standard deviation of the training set channels. The obtained standardized values ​​are retained at the original time step and channel position to obtain a standardized time-series runtime data matrix. .

[0076] Among them, the standardized time-series running data matrix The size is It is a time-series running data matrix that reduces the differences in channel dimensions and numerical ranges among different running data by utilizing the channel statistics of normal training samples. Furthermore, the standardized time-series running data matrix... The relative changes of each running data channel within 60 time steps are preserved, enabling subsequent multi-scale temporal convolutions to extract dynamic change features at a unified numerical scale.

[0077] 5> Transfer the physical feature vector and standardized time-series running data matrix Together, they serve as the output of S2. Among them, the standardized time-series runtime data matrix... Used to input multi-scale temporal convolution modules, physical feature vectors Used to generate gate weights that correspond one-to-one with multi-scale temporal feature channels.

[0078] It should be noted that the standardized time-series running data matrix Retain dynamic change information over a continuous 60 seconds, physical feature vector Highlighting sample-level features with clear physical meaning, such as the widening of temperature difference in the cooling circuit, vibration and shock, and fluctuations in current conditions, while retaining both continuous time-series data and sample-level physical features, can avoid the loss of dynamic information caused by using only manual statistical features, and can also avoid the problem that subtle degradation patterns are difficult for the model to actively focus on when only time-series running data is used.

[0079] S3, Construction and Supervised Training of Multi-Scale Temporal Convolutional Networks Standardized time-series running data matrix This data includes changes from four operational data channels over a continuous 60 seconds. Different equipment failures correspond to discriminative time patterns with varying durations and scales of change. Using a single convolutional kernel is insufficient to simultaneously extract short-term shocks, moderate temporal fluctuations, and longer-term trends. Furthermore, the multi-scale convolutional feature channels for different operational samples have varying degrees of importance. Therefore, directly applying the physical feature vectors before the classification layer... When concatenating with convolutional features, it is difficult to adjust the contribution of each feature channel using physical correlation information before summarizing temporal features.

[0080] This invention utilizes three parallel one-dimensional dilated convolution branches to extract operating patterns at different time scales, and leverages physical feature vectors. Gated vectors are generated, and channel-by-channel recalibration is performed on the multi-scale time-series feature maps to ensure that the network preferentially retains feature responses that match cooling circuit degradation, vibration shock, and current conditions. The specific steps are as follows:

[0081] S301, Multi-scale dilated convolutional temporal coding The standardized temporal data matrix is ​​converted into a convolutional input matrix. Complementary multi-scale temporal features are extracted through three parallel one-dimensional dilated convolutional branches with different kernel lengths and dilation rates. The output feature maps of each branch are then concatenated along the channel dimension to form a multi-scale temporal feature map containing rich temporal pattern information, such as... Figure 4 As shown, the specific steps are as follows: 1> Standardize the time-series running data matrix Transpose the matrix to obtain the convolution input matrix. Convolution input matrix The size is The first dimension corresponds to 4 running data channels, and the second dimension corresponds to 60 time steps.

[0082] It should be noted that the transpose operation only changes the standardized time-series running data matrix. The dimensional arrangement does not change any standardized value or its corresponding sampling time and running data channel, so that the subsequent one-dimensional dilated convolution can slide along the time dimension, and the four running data channels are used together as the input features of each time position.

[0083] 2> Input the convolution matrix Simultaneously inputting three parallel one-dimensional dilated convolution branches. Specifically, the convolution branch indices are denoted as... , The value range is from 1 to 3.

[0084] The first one-dimensional dilated convolution branch has a kernel length of 3 and an inflation rate of 1; the second one-dimensional dilated convolution branch has a kernel length of 5 and an inflation rate of 2; and the third one-dimensional dilated convolution branch has a kernel length of 7 and an inflation rate of 3. All three one-dimensional dilated convolution branches have a stride of 1, and each branch outputs 16 convolutional feature channels.

[0085] It should be noted that the first one-dimensional dilated convolution branch extracts short-term changes between adjacent time steps through continuous time positions, mainly used to respond to vibration shocks or rapid changes in operational data; the second one-dimensional dilated convolution branch extracts fluctuation patterns within a medium time range through interval sampling with an expansion rate of 2; and the third one-dimensional dilated convolution branch extracts trends with a larger span through interval sampling with an expansion rate of 3, mainly used to respond to slower dynamic changes such as a continuous increase in temperature difference.

[0086] It should also be noted that dilated convolution expands the temporal receptive field by setting sampling intervals between adjacent convolutional kernel weights, without increasing the number of convolutional kernel weights. The three one-dimensional dilated convolution branches use different convolutional kernel lengths and dilation rates, enabling the network to simultaneously obtain temporal features at short, medium, and long time scales at the same layer.

[0087] 3> Set the time dimension padding according to the kernel length and dilation rate of each one-dimensional dilated convolution branch, so that the output time length of the three one-dimensional dilated convolution branches is kept at 60.

[0088] In specific implementation, the first The effective kernel length of each one-dimensional dilated convolution branch is: ,in, Indicates the first The kernel length of a one-dimensional dilated convolution branch. Indicates the first The dilation rate of each one-dimensional dilated convolution branch. Since the effective kernel lengths of the three one-dimensional dilated convolution branches are all odd numbers, they can be filled on both sides of the time dimension. One position, that is ,in, Indicates the first The number of fills on one side of each one-dimensional dilated convolution branch in the time dimension. The first one-dimensional dilated convolution branch has a kernel length of 3 and an inflation rate of 1, corresponding to a fill of 1 on one side; the second one-dimensional dilated convolution branch has a kernel length of 5 and an inflation rate of 2, corresponding to a fill of 4 on one side; the third one-dimensional dilated convolution branch has a kernel length of 7 and an inflation rate of 3, corresponding to a fill of 9 on one side.

[0089] It should be noted that the aforementioned padding amount matches the effective kernel length of the three one-dimensional dilated convolution branches, ensuring that each of the three branches outputs 60 time positions with a stride of 1. This avoids the inability to directly concatenate along the feature channel dimension due to differences in output time lengths. Furthermore, the boundary positions can be filled using the conventional zero-padding method found in convolutional networks.

[0090] 4> Perform batch normalization and ReLU activation sequentially on the convolution output of each one-dimensional dilated convolution branch to obtain the corresponding one-dimensional dilated convolution branch feature map. Wherein, the... The feature map of a one-dimensional dilated convolution branch is denoted as... , The size is .

[0091] It should be noted that batch normalization is used to stabilize the numerical distribution of convolutional features in different training batches, and ReLU activation is used to introduce non-linear expressive power; the kernel parameters and batch normalization parameters of the three one-dimensional dilated convolution branches are trained independently, so that different one-dimensional dilated convolution branches can learn the operating mode at different time scales.

[0092] Furthermore, the feature maps of the three one-dimensional dilated convolution branches are concatenated along the feature channel dimension to obtain a multi-scale temporal feature map. , represented as: ; in, This indicates that splicing is performed along the feature channel dimension; This represents the feature map of the first one-dimensional dilated convolution branch. This represents the feature map of the second one-dimensional dilated convolution branch. This represents the feature map of the third one-dimensional dilated convolution branch.

[0093] It should be noted that multi-scale temporal feature maps It simultaneously includes time series features corresponding to short-term shocks, moderate time fluctuations, and longer-term trends, and also features multi-scale time series feature maps. The size is , of which 48 is obtained by splicing together the 16 convolutional feature channels output by each of the three one-dimensional dilated convolutional branches; This indicates the time length after convolution. When using a convolution stride of 1 and the padding method described above, It is 60.

[0094] It should also be noted that the three one-dimensional dilated convolution branches do not repeatedly process the input data, but rather extract complementary temporal information by using different temporal receptive fields. This allows vibration impact, current fluctuations, and slow temperature difference to form more obvious feature responses in the one-dimensional dilated convolution branches that match their durations.

[0095] S302, Access Control and Fault Classification Multi-scale temporal feature maps The 48 convolutional feature channels are generated by different one-dimensional dilated convolutional branches and different convolutional kernels. Different convolutional feature channels contribute differently to the fault identification of the current running sample.

[0096] This invention utilizes physical feature vectors Generate gate weights that correspond one-to-one with the 48 convolutional feature channels, enabling physical information such as cooling circuit temperature degradation, vibration shock, and current fluctuations to directly participate in the multi-scale temporal feature map. Channel recalibration, such as Figure 5 As shown, the specific steps are as follows: 1> 8-dimensional physical feature vector The input is a gated fully connected layer, which performs a linear combination on the eight physical features to obtain the gated intermediate vector. ,Right now .

[0097] in, It refers to the first The gated intermediate vector of each running sample has a dimension of 48; This represents the weight matrix of the gated fully connected layer, with size . , are trainable parameters; This represents the bias vector of the gated fully connected layer, with a dimension of 48, and is a trainable parameter.

[0098] It should be noted that the gated full-connection layer will reduce the temperature difference degradation index of the cooling circuit. Number of vibration and impact cycles Vibration kurtosis Current standard deviation The four running data channels are mapped horizontally to 48 convolutional feature channels, ensuring that each convolutional feature channel receives an adjustment amount determined by the physical state of the current running sample. Furthermore, the gated fully connected layer does not have an activation function to preserve the positive and negative changes in the linear mapping result; the adjustment amount is then converted into gate weights by the subsequent Sigmoid function.

[0099] 2> Gated intermediate vector Perform an element-wise sigmoid mapping to obtain the gated vector. ,Right now .in, It refers to the first The gating vector of each running sample has a dimension of 48; Represents the Sigmoid function; gate vector The The elements are denoted as , This represents the convolutional feature channel index, ranging from 1 to 48. Each gate weight... All are located between 0 and 1, and are associated with multi-scale temporal feature maps. The first in Each convolutional feature channel corresponds to one.

[0100] It should be noted that when the gating weight When the value is close to 1, the dominant response of the corresponding convolutional feature channel is retained; when the gate weights are .... When the value is close to 0, the response of the corresponding convolutional feature channel is suppressed. This is due to the gate vector... The physical feature vector of the currently running sample Different running samples can obtain different convolutional feature channel weights.

[0101] In one embodiment, for example, when the cooling circuit temperature difference degradation index Larger and average oil temperature At relatively high levels, gated fully connected layers can learn to improve the weights of convolutional feature channels that are related to longer-term trends during supervised training; when the number of vibration shocks... and vibration kurtosis When the signal is large, the gated fully connected layer can learn to improve the weights of convolutional feature channels related to short-term shocks. The above correspondence is automatically determined by the supervised training process based on the real-state category labels, without requiring manual specification of fixed connections between a specific physical feature and a specific convolutional feature channel.

[0102] 3> Gating vector Copy along the time dimension to Each time location, and its correlation with multi-scale temporal feature maps. Perform channel-by-channel multiplication to obtain the gated weighted feature map. ,Right now ,and, , .

[0103] in, It refers to the first The gated weighted feature map of the running samples, with a size of ; Represents the gated weighted feature map In the The convolutional feature channel, the first The value at each time position; Representing multi-scale temporal feature maps The value at the corresponding position; Represents the gate vector Middle and the first The gating weights corresponding to each convolutional feature channel.

[0104] It should be noted that the same gating weights are used for the same convolutional feature channel at all time points. This adjusts the overall contribution of the convolutional feature channel over the entire time range without altering the original temporal sequence and relative relationships within the channel. Furthermore, this process involves recalibrating the convolutional feature channel, rather than independently selecting individual time points.

[0105] 4> Gated weighted feature maps along the time dimension Perform global average pooling, dividing each convolutional feature channel into... The responses at each time point are summarized into a single value to obtain the classification feature vector. Classification feature vector The element for .

[0106] in, It refers to the first The classification feature vector of each sample is 48 in dimension. Refers to classification feature vector The The element is used to represent the first element. The average response of each convolutional feature channel over the full time range after physical feature gating.

[0107] It should be noted that global average pooling aggregates feature responses at different length positions into a fixed-length vector, enabling the classification fully connected layer to receive inputs of fixed dimensions while retaining the contribution of convolutional feature channels after physical feature gating.

[0108] 5> Classification feature vector The input is a fully connected classification layer, which performs a linear combination on the 48 classification features to obtain the classification score vector. ,Right now Then, the classification score vector Performing Softmax mapping yields the fault probability vector. ,Right now .

[0109] in, Indicates the first The classification score vector of each running sample has a dimension of 4; This represents the weight matrix of the fully connected classification layer, with size . These are trainable parameters; This represents the bias vector of the fully connected classification layer, which has a dimension of 4 and is a trainable parameter. Indicates the first The failure probability vector of each running sample, i.e. ;in, Indicates the first The running sample belongs to the first The model output probability of each device operating state category (e.g., Indicates the first (The model output probability of each running sample belonging to a cooling system failure). This represents the category index, with a value ranging from 1 to 4; the sum of the probabilities output by the four models is 1.

[0110] 6> During the supervised training phase, the true state category labels of the training samples are assigned. With the failure probability vector The cross-entropy loss function is input as a common input. Backpropagation is performed based on the cross-entropy loss to update all trainable parameters of the three one-dimensional dilated convolutional branches, the batch normalization layer, the gated fully connected layer, and the classification fully connected layer.

[0111] In the specific implementation, S2, S301, and S302 are executed sequentially for the running samples in each training batch to obtain the corresponding fault probability vector. Then, categorize by actual state. Read the fault probability vector The model output probabilities for the corresponding categories are calculated, and the cross-entropy loss is also calculated. By minimizing the cross-entropy loss, the model output probabilities for the true state categories are gradually increased, while the model output probabilities for other state categories are gradually decreased.

[0112] During training, the temperature difference degradation index of the cooling circuit Number of vibration and impact cycles Vibration kurtosis Other physical features are not used as trainable parameters, but are directly calculated based on each running sample. Furthermore, the gated fully connected layer learns the physical feature vector through supervised training. The mapping relationship between the feature channels and the 48 convolutional feature channels.

[0113] It should be noted that physical feature vectors It does not directly replace the standardized time-series running data matrix It's not just before the classification layer and the classification feature vector. Instead of performing end-cap splicing, it uses a gated vector. Participating in multi-scale temporal feature maps The recalibration of convolutional feature channels enables the network to utilize the cooling circuit temperature difference degradation index, vibration and shock features, and current condition information before summarizing time-series features. This reduces the interference of feature responses inconsistent with the current physical state on the classification results and improves the recognition stability under small sample and weak fault feature conditions.

[0114] It should also be noted that the multi-scale temporal convolutional network consists of three parts: feature extraction, channel gating, and classification. The specific hierarchical structure is shown in the table below:

[0115] In one embodiment, such as Figure 6-8 As shown, the gating weight vectors of samples in different states are compared, specifically the gating weight vectors corresponding to three types of samples: normal state, mechanical failure, and cooling system failure. The horizontal axis of each subgraph represents the index of 48 feature channels, and the vertical axis represents the gating weight value (range 0 to 1). The channel gating weights generated by the physical feature vectors through the gating fully connected layer can dynamically adjust the importance of multi-scale feature channels according to the current physical state of the device, achieving an organic fusion of physical knowledge and deep learning features.

[0116] S4. Status confirmation and early warning decision-making based on the synergy between model output and degradation index Fault probability vector output by a multi-scale temporal convolutional network It reflects the classification results based on time-series patterns, but lacks constraints on the consistency of physical mechanisms. At the same time, the confidence level of a single sample cannot effectively distinguish between persistent degradation and transient perturbations.

[0117] This invention utilizes physical feature vectors Construct a physical matching vector for the fault probability vector. Perform class-by-class consistency correction to obtain the cooperative probability vector. Furthermore, by combining the sliding window trend analysis of the cooling circuit temperature difference degradation index, the judgment confidence level and degradation trend sharpness of the current sample are nonlinearly fused to generate an early warning index. Based on this, a final alarm decision is output to balance the sensitivity of identification and the reliability of alarm in the early stage of fault detection. The specific steps are as follows: S401, Calculation of State Confidence Weighted by Physical Consistency

[0118] Physical feature vector Gated vectors have been used Convolutional feature channel recalibration is involved in multi-scale temporal feature maps, but this function is implicitly learned through supervised training, and the fault probability vector... It does not directly constrain the explicit matching relationship between each fault category and its corresponding physical quantity; when the equipment is in the early stage of failure, the temporal variation pattern is not yet significant, or the temporal patterns of different faults are locally similar, the fault probability vector There may be issues where multiple fault categories have similar probabilities.

[0119] This invention starts from physical feature vectors Extract physical quantities that have a clear physical association with each fault category, construct a physical matching vector, and then apply this to the fault probability vector. A state-by-state category correction is performed to obtain the cooperative probability vector, and the maximum component of the cooperative probability vector is used as the final decision confidence level. The specific steps are as follows: 1> From physical feature vectors Physical quantities with a clear physical correlation to each fault category are selected. Specifically, for mechanical faults, the number of vibration and impact cycles is selected. and vibration kurtosis For cooling system failures, select the cooling circuit temperature difference degradation index. For electrical faults, select the current standard deviation. .

[0120] It should be noted that the physical matching vector The design goal is to measure the degree to which the current physical quantity deviates from the "normal baseline" in order to enhance support for specific fault categories. The normal state is the state in which all physical quantities do not deviate significantly from their normal baseline. It does not have a specific, directional physical anomaly pattern to match. Therefore, no physical bias correction is applied to the normal state, and the corresponding physical matching weight is fixed at 1.

[0121] It should also be noted that this selection method matches the probability of mechanical failure with the number of vibrations and the degree of vibration spikes, matches the probability of cooling system failure with the degree of continuous expansion of temperature difference, and matches the probability of electrical failure with the degree of fluctuation of the effective value of A-phase current, thus avoiding the use of features that lack direct physical correlation to correct the corresponding fault categories.

[0122] 2. Calculate the normal baseline mean and standard deviation of each physical quantity using the normal training samples. Specifically, let the total number of normal training samples be denoted as... Calculate the mean number of vibration and impact events for the normal training samples respectively. and standard deviation of vibration and shock number , mean kurtosis and vibration kurtosis standard deviation Average temperature difference degradation index of cooling circuit and the standard deviation of the cooling circuit temperature difference degradation index Mean of standard deviation of current and current standard deviation benchmark standard deviation .

[0123] Among them, the current standard deviation is the benchmark standard deviation. This refers to the standard deviation of current among normal training samples. The standard deviation is not the current standard deviation of the effective value sequence of phase A current within a single operating sample. .

[0124] It should be noted that the above-mentioned normal baseline statistics are determined once during the model training phase and remain unchanged during the model validation and online deployment phases. Furthermore, the normal training samples should cover multiple load levels and natural fluctuations under normal operating conditions to ensure that the standard deviation of each normal baseline is positive and can reflect the reasonable range of variation in normal operating conditions.

[0125] 3> Regarding the first For each running sample, the number of vibration and impact cycles was calculated. Vibration kurtosis Cooling circuit temperature difference degradation index and current standard deviation The standardized deviation is compared to the corresponding normal baseline, and a biased Sigmoid function is used to map the standardized deviation to physical matching weights in the range of 0 to 1, forming the first... Physical matching vector of the running sample , represented as: ; Among them, physical matching vector It is a 4-column vector; Represents the Sigmoid function; and The physical matching scale factor for mechanical faults is used to control the degree of influence of vibration impact number deviation and vibration kurtosis deviation on the physical matching weight of mechanical faults, respectively. This represents the physical matching scale factor for cooling system faults, used to control the degree of influence of the deviation of the cooling circuit temperature difference degradation index on the weight of the physical matching of cooling system faults. This represents the physical matching scale factor for electrical faults, used to control the degree of influence of the current standard deviation on the physical matching weight of electrical faults.

[0126] In one implementation, , , and All are positive numbers, which can be determined by a grid search of the validation set and remain unchanged after online deployment. , and These represent the physical matching biases corresponding to mechanical faults, cooling system faults, and electrical faults, respectively. They are used to set the basic matching weights for the corresponding fault categories when the physical quantities are close to the normal baseline, thus avoiding excessive suppression of the probability of all fault categories by physical consistency correction.

[0127] It should be noted that physical matching bias , and This is used to set a baseline matching weight to ensure that when the physical quantity deviates from the normal baseline by zero (i.e., is completely within the normal range), the corresponding fault category still has a certain possibility of being identified. This avoids the physical consistency correction from excessively suppressing the fault probability to near zero, thereby completely rejecting the model's timing judgment. In one implementation, the output of the Sigmoid function can be set to 0.5 when the standardized deviation of the physical quantity is 0, meaning that the physical evidence neither supports nor opposes the fault. Based on this, it can be deduced that... The general calculation formula. For example, let and ,but Solving for Similarly, , In another implementation, if a higher or lower base weight is desired, a grid search of the validation set can be used to find the bias value that optimizes the overall performance of the validation set in the vicinity of 0 (e.g., the [-1,1] interval).

[0128] It should also be noted that the physical matching vector The first element in The first element corresponds to the normal state, indicating that no physical penalty or enhancement is imposed on the probability of the normal state; the second element... Corresponding to mechanical failures, when the number of vibration and impact cycles... and vibration kurtosis When the elevation is relative to the normal baseline, Increase; the 3rd element In response to a cooling system failure, when the temperature difference degradation index of the cooling circuit... When the elevation is relative to the normal baseline, Increase; 4th element For electrical faults, when the current standard deviation When the elevation is relative to the normal baseline, Increase.

[0129] 4> Utilizing physical matching vectors For the fault probability vector Perform element-wise weighting, then divide the weighted probability of each state category by the sum of the weighted probabilities of all state categories to obtain the nth... Collaborative probability vector of the running samples , represented as: ; Wherein, the cooperative probability vector The dimension is 4, which is represented as ; This indicates element-wise multiplication; Indicates the first The running sample belongs to the first The model outputs probabilities for each state category; Indicates the first The running sample is in the first The physical matching weights for each state category; the denominator represents the sum of the weighted probabilities of the four state categories, making the cooperative probability vector... The sum of the four components remains 1.

[0130] For example, when the mechanical failure probability and the cooling system failure probability in the failure probability vector are close, but the vibration impact number and vibration kurtosis of the current operating sample are significantly higher than the normal baseline, while the cooling circuit temperature difference degradation index is close to the normal baseline, the physical matching weight corresponding to the mechanical failure is higher than the physical matching weight corresponding to the cooling system failure. The normalized cooperative probability vector will relatively increase the probability proportion of mechanical failure.

[0131] 5> From the cooperative probability vector Select the state category corresponding to the largest component from the list to obtain the first... Final state category of the running sample ,Right now Then, the largest component is used as the final decision confidence level. ,Right now .

[0132] Among them, the final state category The value ranges from 1 to 4, corresponding to normal state, mechanical failure, cooling system failure, and electrical failure, respectively; the final confidence level is determined. Used to characterize the reliability of the final state determination of the currently running sample.

[0133] It should be noted that the physical consistency weighting mechanism does not replace the multi-scale temporal convolutional network. Instead, after the multi-scale temporal convolutional network completes state probability prediction, it uses physical quantities with clear fault mechanism associations to perform consistency correction on the probabilities of each state category. When the network output probability of a certain fault category is high and the corresponding physical anomaly also exists, the relative proportion of that fault category in the cooperative probability vector is preserved or increased; when the network output probability of a certain fault category is high but the corresponding physical quantity is still close to the normal range, the relative proportion of that fault category in the cooperative probability vector is suppressed, thereby reducing misclassification caused by accidental similarity of temporal signals.

[0134] S402, Dynamic early warning decision based on the sharpness of the degradation trend Final judgment confidence level of a single running sample This only reflects the status assessment results within the current 60-second time window. Hydropower station sites experience non-fault transients such as short-term load adjustments, sensor spikes, and short-term temperature changes in the cooling medium. If alarms are triggered solely based on high-confidence status identification results from a single operational sample, false alarms are likely. The actual equipment degradation process typically has a certain duration, especially the temperature difference degradation index in the cooling circuit. As the cooling system slowly deteriorates, it will show a continuous upward trend.

[0135] This invention measures the temperature difference degradation index of cooling circuits for multiple consecutive operating samples. Perform exponentially weighted moving average and trend smoothing calculations to determine the final confidence level of the current running sample. The positive trend of the cooling circuit temperature difference degradation index is nonlinearly fused to obtain the early warning index, and an alarm decision is output based on the early warning index. The specific steps are as follows:

[0136] 1> The maintenance length during online operation is A degradation trend sliding window, used to store the most recent degradation trends. Temperature difference degradation index of smooth cooling loop corresponding to each operating sample ,in, The possible value is 5. The index of the currently running sample is... At that time, the degradation trend sliding window contains in sequence... .

[0137] Furthermore, an exponentially weighted moving average is applied to the cooling circuit temperature difference degradation index sequence to obtain a smoothed cooling circuit temperature difference degradation index. The recursive process is represented as: ;

[0138] in, Indicates the first Temperature difference degradation index of the smooth cooling loop corresponding to each operating sample; Indicates the first Temperature difference degradation index of the smooth cooling loop corresponding to each operating sample; Represents the exponentially weighted moving average smoothing factor, satisfying This is used to control the degree to which historical samples retain the current smoothing result, and can be set to 0.8; the initial value corresponding to the first sample is set to... .

[0139] In one implementation, when the exponentially weighted moving average smoothing factor... When the value is 0.8, the temperature difference degradation index of the smoothed cooling circuit corresponding to the previous running sample accounts for 0.8 of the current smoothing result, and the temperature difference degradation index of the cooling circuit of the current running sample accounts for 0.2 of the current smoothing result, thereby reducing the impact of abnormal temperature jumps in a single running sample on trend judgment.

[0140] 2> In the degradation trend sliding window, select the most recent The running samples are arranged in ascending order of time, and a time index is set within the window. Among them, the time index within the window. Corresponding to the oldest Running samples, time index within the window Corresponding to the latest number The first running sample. Therefore, the first sample in the window... The temperature difference degradation index of the smooth cooling circuit corresponding to each location is: .

[0141] Furthermore, to ensure that newer running samples have a greater impact on trend calculation, the first sample within the window... Each position is configured with an exponential decay weight. The weight of the currently running sample is 1, and the weight of the oldest running sample (i.e., the sample with the earliest sampling time and the furthest time from the current moment in the sliding window) is 1. .

[0142] 3> Calculate the weighted average time index based on the in-window time index, exponential decay weight, and smoothing cooling loop temperature difference degradation index. The weighted average of the temperature difference degradation index of the smooth cooling loop within the window. , represented as: ; ; Among them, weighted average time index Used to characterize the center of weight of the exponentially decaying weights on the time axis; weighted average. The weighted baseline level of the temperature difference degradation index of the smooth cooling circuit within the sliding window used to characterize the current degradation trend.

[0143] 4> Utilize the in-window time index relative to the weighted average time index The deviation, and the temperature difference degradation index of the smoothing cooling circuit relative to the weighted average. The bias is used to perform weighted linear regression to obtain the first... Smooth deterioration rate of change corresponding to each running sample , represented as: ; Among them, the smoothing rate of change of deterioration Used to characterize recent The weighted average change direction and rate of change of the temperature difference degradation index of the smooth cooling circuit in the sample, in units of cooling circuit temperature difference degradation index per second.

[0144] It should be noted that as the cooling circuit continues to deteriorate, the temperature difference deterioration index of the smoothing cooling circuit within the window gradually increases from old to new, and the smoothing deterioration change rate... It is a positive value; when the temperature difference degradation trend in the cooling circuit stops, the smoothed degradation rate is... Approaching 0; when the degree of temperature difference degradation in the cooling circuit decreases, the rate of change of degradation is smoothed out. It is a negative value.

[0145] It should also be noted that the time index within the window increases from oldest to newest to ensure a smooth rate of degradation. The positive and negative directions are consistent with the actual deterioration direction of the cooling circuit, avoiding errors in the trend sign caused by the opposite direction of time arrangement.

[0146] 5> Smooth the rate of change of degradation Divide by the standard deviation of the cooling circuit temperature difference degradation index of the normal training samples This yields the dimensionless rate of change relative to the normal fluctuation range, and then uses exponential mapping to correlate the positive deterioration trend with the final judgment confidence level of the current operating sample. By merging, we obtain the first... Early warning index of the sample of running , represented as: ; Among them, the early warning index Used to comprehensively characterize the reliability of the current state determination and the recent trend of temperature difference degradation in the cooling circuit; The sharpness coefficient of the degradation trend is used to adjust the impact of a positive degradation trend on the early warning index. The degree of magnification, and This can be determined through the validation set; The standard deviation of the cooling circuit temperature difference degradation index represents the normal training samples and is used to characterize the natural fluctuation range of the cooling circuit temperature difference degradation index under normal operating conditions. This indicates that the early warning index is used to calculate the stability constant, which is used to avoid... When the value is 0, division by zero occurs, which can be taken as... ; This indicates that only the smoothed rate of change of degradation is retained. The positive part.

[0147] It should be noted that when smoothing the rate of change of degradation When the index is not greater than 0, the index term is 1, indicating a warning index. Equals the final judgment confidence level It does not amplify stable or declining trends; when smoothing the rate of change of deterioration. When it is greater than 0, the warning index is It increases with the positive deterioration trend, enabling a stronger early warning response to continuously deteriorating cooling system conditions compared to a single high-confidence anomaly.

[0148] It should also be noted that the smoothing rate of change of degradation Deterioration index derived from cooling circuit temperature difference Therefore, trend index mapping is mainly used to enhance the early warning intensity corresponding to the continuous deterioration of the cooling system. For mechanical or electrical faults, when the cooling circuit temperature difference deterioration index does not form a positive trend, the early warning index... The final determination of confidence level still depends on the outcome. The decision will not be amplified simply because there is no trend of cooling degradation.

[0149] 6> Increase the warning index With warning threshold Comparison. Specifically, when the warning index... Not less than the warning threshold When the fault alarm signal is output, the final status category will be determined. As a fault alarm category; when the warning index Less than the warning threshold At this time, no fault alarm signal is output; only the final state category of the currently running sample is recorded. Final determination of confidence level and early warning index .

[0150] Warning threshold The sharpness coefficient of the degradation trend is determined using a validation set after the multi-scale temporal convolutional network has been trained. In one implementation, the false alarm rate of the normal class validation samples does not exceed a preset level as a constraint. and warning threshold A joint selection was made.

[0151] Furthermore, the average temperature difference degradation index of the cooling circuit corresponding to the normal training samples can be used as a basis. and the standard deviation of the cooling circuit temperature difference degradation index Determine the auxiliary judgment threshold ,Right now Among them, the auxiliary judgment threshold It is used to assist in determining whether the current cooling circuit temperature difference degradation index is significantly higher than the normal fluctuation range, but it is not the sole triggering condition for a fault alarm signal. For example, the auxiliary judgment threshold. As supplementary reference information for early warning, when the early warning index Exceeding the threshold but If not significant, if the current Exceed If so, a 'temperature difference level abnormal' marker can be added to the alarm log for maintenance personnel to refer to.

[0152] It should be noted that this step nonlinearly fuses the final judgment confidence of a single operating sample with the cooling circuit temperature difference degradation trend in continuous operating samples, so that the fault warning considers both the recognition credibility of the current state mode and whether the cooling circuit degradation process shows continuous deterioration characteristics, thereby reducing false alarms caused by short-term load changes, temperature jumps or sensor spikes.

[0153] S5. Online Status Identification and Early Warning of Secondary Control Equipment in Hydropower Stations After completing supervised training of the multi-scale temporal convolutional network, reference values ​​for the current standard deviation are determined and stored using normal class training samples. The average values ​​of the training set channels corresponding to the four running data channels. and the standard deviation of the training set channels Average number of vibration and impact cycles and standard deviation of vibration and shock number , mean kurtosis and vibration kurtosis standard deviation Average temperature difference degradation index of cooling circuit and the standard deviation of the cooling circuit temperature difference degradation index Mean of standard deviation of current and current standard deviation benchmark standard deviation .

[0154] Furthermore, the physical matching scale factor is determined through the validation set. , , and Physical matching bias , and Deterioration trend sliding window length Exponentially weighted moving average smoothing factor Deterioration trend sharpness coefficient and warning threshold Auxiliary judgment threshold according to The aforementioned fixed parameters, along with the trainable parameters of the multi-scale temporal convolutional network, are then deployed to the condition monitoring system of the secondary industrial control equipment in the hydropower station.

[0155] During online operation, the hydropower station's secondary industrial control equipment status monitoring system continuously collects the effective value of vibration acceleration, the effective value of phase A current, oil temperature, and cooling water temperature according to the same operating data channel configuration and 1-second sampling interval as during the training phase, and constructs one sample to be tested with 60 consecutive sampling time steps.

[0156] For each sample to be tested, S2 is executed first to reconstruct the sample into a time-series running data matrix and calculate the cooling loop temperature difference degradation index. and vibration-current coupling eigenvectors Further, physical feature vectors are obtained. and standardized time-series running data matrix The standardization process continues to utilize the channel mean of the training set saved during the training phase. and the standard deviation of the training set channels .

[0157] Then execute S3 to standardize the time-series running data matrix. Input three parallel one-dimensional dilated convolution branches and utilize physical feature vectors. Generate gate vector After convolutional feature channel recalibration, global average pooling, classification fully connected layer, and Softmax mapping, the fault probability vector is obtained. .

[0158] Further executing S401, the physical matching vector is calculated using the normal baseline statistics saved during the training phase and the physical matching parameters determined during the validation phase. and using physical matching vectors For the fault probability vector Perform state-by-state category correction to obtain the cooperative probability vector. The collaborative probability vector The state category corresponding to the largest component is used as the final state category. The largest component is used as the final confidence level. .

[0159] Starting with the first sample to be tested, the temperature difference degradation index of the smoothed cooling circuit is calculated according to the exponentially weighted moving average recursive relationship. And the current cooling circuit temperature difference degradation index and its corresponding smooth cooling circuit temperature difference degradation index Add length is The deterioration trend sliding window. When the cumulative number of samples to be tested is insufficient... At that time, the smoothing rate of change of degradation is not calculated. and early warning index Only record the final state category and final judgment confidence level When the cumulative number of samples to be tested reaches Then, the smoothing rate of change of degradation is calculated according to S402. and early warning index .

[0160] When the warning index Not less than the warning threshold At that time, the secondary control equipment status monitoring system of the hydropower station outputs a fault alarm signal and the final status category. Final determination of confidence level Current cooling circuit temperature difference degradation index and smoothing degradation rate When the warning index Less than the warning threshold At that time, the secondary industrial control equipment status monitoring system of the hydropower station does not output fault alarm signals, but only records the corresponding status monitoring log.

[0161] It should be noted that the current standard deviation reference value determined during the training phase must continue to be used during the online status recognition process. Training set channel mean Training set channel standard deviation The normal baseline mean and standard deviation of normal physical quantities must not be recalculated based on a single sample to be tested or a sample to be tested in the near future. Physical matching scale factor, physical matching bias, and degradation trend sliding window length are also included. Exponentially weighted moving average smoothing factor Deterioration trend sharpness coefficient Warning threshold and auxiliary judgment threshold After deployment, it should be kept fixed to ensure that the numerical benchmarks and early warning conditions of the status identification results corresponding to different time windows remain consistent.

[0162] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for multi-source data fusion monitoring of the operating status of secondary control equipment in hydropower stations, characterized in that, Includes the following steps: S1. Collect multi-channel raw operating data during the continuous operation of the secondary industrial control equipment of the hydropower station. The multi-channel raw operating data includes the effective value of vibration acceleration, the effective value of phase A current, oil temperature and cooling water temperature. Multiple consecutive sampling time steps constitute an operating sample, and a real state category label is configured for each operating sample. S2. Reconstruct the original operating data into a time-series operating data matrix, construct the cooling circuit temperature difference degradation index and vibration-current coupling characteristics, and concatenate the cooling circuit temperature difference degradation index, vibration-current coupling characteristics and the basic statistics of each operating data channel into a physical feature vector. Then, use the channel statistics of the normal class training samples to perform channel-level standardization on the time-series operating data matrix to obtain a standardized time-series operating data matrix. The cooling circuit temperature difference degradation index is used to quantify the degree of continuous expansion of the temperature difference between oil temperature and cooling water temperature. The vibration-current coupling characteristics simultaneously include the number of vibration impacts, the degree of vibration peaks, and the level of current fluctuation. S3. Construct a multi-scale temporal convolutional network, use multiple parallel one-dimensional dilated convolutional branches to extract temporal features at different time scales from the standardized temporal running data matrix to obtain a multi-scale temporal feature map, and use the physical feature vector to generate a gating vector, perform channel-by-channel recalibration on the multi-scale temporal feature map, and then output a fault probability vector through a classification layer. S4. Construct a physical matching vector using the physical feature vector, perform category-by-category consistency correction on the fault probability vector to obtain a cooperative probability vector, and determine the final state category and final judgment confidence level based on the cooperative probability vector; simultaneously, perform sliding window trend analysis on the cooling circuit temperature difference deterioration index of multiple consecutive operating samples, fuse the final judgment confidence level with the positive change trend of the cooling circuit temperature difference deterioration index to obtain a warning index, and output an alarm decision based on the warning index.

2. The multi-source data fusion monitoring method for the operating status of secondary control equipment in hydropower stations according to claim 1, characterized in that, The methods for constructing the cooling circuit temperature difference degradation index described in S2 include: Extract the oil temperature sequence and the coolant temperature sequence from the time-series running data matrix, calculate the difference between the oil temperature and the coolant temperature at the same time step, and obtain the temperature difference sequence. The temperature difference sequence is subjected to least squares linear fitting to obtain the linear fitting slope and linear fitting intercept, and the temperature difference fitting error is calculated. The positive portion of the linear fitting slope is retained, and the positive change rate is weighted using the temperature difference fitting error to obtain the cooling circuit temperature difference degradation index.

3. The multi-source data fusion monitoring method for the operating status of secondary control equipment in hydropower stations according to claim 1, characterized in that, The construction methods of the vibration-current coupling characteristics described in S2 include: The effective value sequence of vibration acceleration and the effective value sequence of phase A current are extracted from the time-series running data matrix, and the vibration mean, vibration standard deviation, current mean and current standard deviation are calculated respectively. Calculate the current standard deviation corresponding to each normal class training sample, and perform an arithmetic mean on the current standard deviations of all normal class training samples to obtain a reference value for the current standard deviation. Based on the deviation between the current standard deviation of the current sample and the reference value of the current standard deviation, an adaptive threshold coefficient is generated using the hyperbolic tangent function; The vibration impact detection threshold is determined based on the vibration mean, vibration standard deviation and adaptive threshold coefficient. The effective value of vibration acceleration is compared with the vibration impact detection threshold step by step, and the number of vibration impacts is counted. Calculate the vibration kurtosis, and concatenate the vibration impact number, the vibration kurtosis, and the current standard deviation into a vibration-current coupling feature vector.

4. The multi-source data fusion monitoring method for the operating status of secondary control equipment in a hydropower station according to claim 1, characterized in that, The physical feature vector includes: cooling circuit temperature difference degradation index, vibration impact number, vibration kurtosis, current standard deviation, average oil temperature, average cooling water temperature, average vibration, and average current, and the features in the physical feature vector are arranged in a fixed order.

5. The multi-source data fusion monitoring method for the operating status of secondary control equipment in a hydropower station according to claim 1, characterized in that, In S3, a multi-scale temporal convolutional network is constructed. Multiple parallel one-dimensional dilated convolutional branches are used to extract temporal features at different time scales from the standardized temporal data matrix, resulting in a multi-scale temporal feature map. Specifically, this includes: The standardized time-series running data matrix is ​​transposed to obtain the convolution input matrix; The convolution input matrix is ​​simultaneously input into three parallel one-dimensional dilated convolution branches; The time dimension padding is set according to the kernel length and dilation rate of each one-dimensional dilated convolution branch to keep the output time length of the three one-dimensional dilated convolution branches consistent. Batch normalization and ReLU activation are performed sequentially on the convolution output of each one-dimensional dilated convolution branch to obtain the corresponding one-dimensional dilated convolution branch feature map.

6. The multi-source data fusion monitoring method for the operating status of secondary control equipment in a hydropower station according to claim 1, characterized in that, In S3, the physical feature vector is used to generate a gate vector, and channel-by-channel recalibration is performed on the multi-scale temporal feature map. Specifically, this includes: The physical feature vector is input into the gated fully connected layer to obtain the gated intermediate vector; Sigmoid mapping is performed element-wise on the gated intermediate vector to obtain a gated vector, wherein the dimension of the gated vector is the same as the number of channels of the multi-scale temporal feature map; The gated vector is copied along the time dimension and multiplied channel-by-channel with the multi-scale temporal feature map to obtain a gated weighted feature map.

7. The multi-source data fusion monitoring method for the operating status of secondary control equipment in a hydropower station according to claim 1, characterized in that, The physical matching vectors described in S4 are constructed in the following ways: Physical quantities that are physically related to each fault category are selected from the physical feature vector. For mechanical faults, the number of vibration impacts and vibration kurtosis are selected; for cooling system faults, the cooling circuit temperature difference deterioration index is selected; and for electrical faults, the current standard deviation is selected. The normal baseline mean and normal baseline standard deviation of each physical quantity were statistically analyzed using normal training samples. Calculate the standardized deviation of each physical quantity of the current running sample relative to the corresponding normal baseline, and use the biased Sigmoid function to map the standardized deviation to physical matching weights in the range of 0 to 1 to form the physical matching vector.

8. The method for multi-source data fusion monitoring of the operating status of secondary control equipment in hydropower stations according to claim 1, characterized in that, In S4, the failure probability vector is corrected for consistency across categories to obtain the cooperative probability vector, specifically including: The physical matching vector is used to perform element-wise weighting on the fault probability vector; The cooperative probability vector is obtained by dividing the weighted probability of each state category by the sum of the weighted probabilities of all state categories.

9. The method for multi-source data fusion monitoring of the operating status of secondary control equipment in hydropower stations according to claim 1, characterized in that, The sliding window trend analysis described in S4 includes: An exponentially weighted moving average is applied to the cooling circuit temperature difference degradation index sequence to obtain a smoothed cooling circuit temperature difference degradation index. Within the deterioration trend sliding window, weighted linear regression is performed using exponential decay weights to calculate the smoothed deterioration rate of change. The early warning index is obtained by dividing the smoothed degradation rate by the standard deviation of the cooling circuit temperature difference degradation index of the normal training samples, and then multiplying it by the final judgment confidence level after index mapping.

10. The multi-source data fusion monitoring method for the operating status of secondary control equipment in a hydropower station according to any one of claims 1 to 9, characterized in that, It also includes S5: After completing the training of the multi-scale temporal convolutional network and determining various fixed parameters, the trainable parameters of the multi-scale temporal convolutional network and the fixed parameters determined by the normal class training samples are deployed to the status monitoring system of the secondary industrial control equipment of the hydropower station; during online operation, multi-channel operation data is continuously collected to form the sample to be detected, and S2, S3 and S4 are executed in sequence. When the warning index is not less than the preset warning threshold, the fault alarm signal and the final status category are output; otherwise, only the status monitoring log is recorded.

Images