Switching power supply health degree evaluation method based on data analysis

Abstract

The invention relates to the technical field of operation state evaluation, in particular to a switching power supply health degree evaluation method based on data analysis, which comprises the following steps of: establishing operation load correlation characteristics through mean value and range calculation, extracting continuous increasing paragraphs in a ratio sequence by using a difference mode, and calculating the health degree of a switching power supply; when the length of the interval exceeds a threshold value, a degradation starting point is recorded, a data segment is input into a long short-term memory network for time sequence modeling, offset is accumulated point by point in the modeling process and is compared with the threshold value in sequence, an abnormal path range is output, a degradation trend judgment result is generated, and a heat dissipation unit temperature rise and magnetic element jitter sequence is divided into four types of states; adjacent state changes are counted one by one, a proportion matrix is generated, convergence distribution is obtained through a Markov model through power accumulation, serious degradation and failure state distribution values are compared with a threshold value point by point, a critical interval is marked when the threshold value is exceeded, and life prediction distribution is formed; and life prediction can be combined with multi-dimensional state changes instead of depending on a single index.

Description

Technical Field

[0001] This invention relates to the field of operational status assessment technology, and in particular to a method for assessing the health of switching power supplies based on data analysis. Background Technology

[0002] The field of operational status assessment technology focuses on the operational status performance of power electronic devices. Specifically, it includes the quantitative monitoring of electrical parameters, electromagnetic characteristics, thermal characteristics, and load adaptability. Through data processing methods, it establishes performance evaluation models to identify potential performance degradation and failure trends of power devices, thereby providing a basis for system maintenance, reliability design, and life prediction.

[0003] A data-driven assessment method for the health of switching power supplies addresses the issues of efficiency degradation, output voltage fluctuations, and increased power loss that often occur during long-term operation. This data-driven assessment scheme aims to quantify the degree of performance degradation of the switching power supply by collecting and analyzing operating data such as voltage, current, temperature rise, switching frequency, and harmonic components. This enables accurate diagnosis of the switching power supply's operating status, reduces the risk of unplanned downtime, and extends equipment lifespan.

[0004] Existing technologies for operational status assessment primarily rely on single-dimensional quantitative monitoring of electrical parameters, thermal characteristics, and electromagnetic characteristics, and perform trend analysis through conventional data processing methods. They lack the ability to capture the complex coupling relationship between voltage, current, and temperature rise at the sequence level, making it difficult to reveal potential degradation patterns. Furthermore, existing methods are not sensitive enough to early degradation initiation points, resulting in degradation paths being identified only when fault signs are obvious. Estimation using empirical models and rates of change based on single parameters fails to consider the state transition characteristics of heat dissipation units and magnetic components during operation, making the results prone to deviating from the actual lifespan evolution. Monitoring methods based on global statistics are insufficient in identifying potential faults, which may lead to delays in maintenance decisions, increase the risk of unplanned downtime, and weaken the operational reliability and continuity of power electronic devices. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a data analysis-based method for assessing the health of switching power supplies.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a data analysis-based method for assessing the health of a switching power supply, comprising the following steps: S1: Based on voltage fluctuation sequence, current fluctuation sequence, and power device operating characteristics, time alignment is performed and instantaneous power value is calculated. Then, it is divided point by point with the temperature rise sequence to generate a ratio sequence. The mean and range of the ratio sequence are calculated to establish the operating load correlation characteristics. S2: Based on the operating load correlation characteristics, calculate the difference segment by segment in the ratio sequence and identify the rising interval. When the interval exceeds the threshold, mark the degradation start point and input it into the long short-term memory network. Extract the cumulative offset of the data segment and compare it with the threshold point by point. Mark the abnormal path in the continuous exceeding interval and generate the degradation trend judgment result. S3: Based on the degradation trend determination result, extract the filter capacitor ripple and temperature rise sequence in the abnormal interval and compare them to generate a difference sequence and calculate the sum of squares to form relevant values. Construct a matrix to filter out the threshold terms and obtain the degradation path feature set. S4: Based on the set of degradation path features, the temperature rise of the heat dissipation unit and the jitter of the magnetic element are assigned to the state set and divided into four categories. The transition is recorded and a ratio matrix is ​​generated. The state transition law is defined by the Markov model. The distribution is obtained by power calculation. When the severe degradation and failure exceed the threshold, the interval is marked as critical to obtain the lifetime prediction distribution result. S5: Based on the lifespan prediction distribution results, superimpose abnormal paths of the control loop in the critical interval, calculate the overlap ratio and compare it with the threshold. If the ratio exceeds the limit, generate a trigger flag and obtain the health assessment conclusion.

[0007] As a further aspect of the present invention, the operating load correlation features include the power ratio mean, power ratio range, and power fluctuation amplitude; the degradation trend determination results include the degradation start point position, cumulative offset value, and abnormal path range; the degradation path feature set includes ripple threshold point, temperature rise threshold point, and matrix anomaly item; the lifetime prediction distribution results include state distribution probability, lifetime critical interval, and remaining lifetime range; and the health assessment conclusions include lifetime critical point, abnormal overlap ratio, and maintenance trigger marker.

[0008] As a further aspect of the present invention, the specific steps for generating the operational load correlation features are as follows: Based on voltage fluctuation sequence, current fluctuation sequence, and power device operating characteristics, voltage and current are multiplied point by point to obtain instantaneous power sequence, and instantaneous power is divided point by temperature rise to form ratio sequence, thus generating ratio sequence result; Based on the ratio sequence results, the mean is obtained by summing the values ​​at each point, and the range is obtained by differentiating the valley and peak values. The combination of the mean and range is recorded as a parameter set to establish the operating load correlation characteristics.

[0009] As a further aspect of the present invention, the specific steps for generating the degradation trend determination result are as follows: Based on the aforementioned operating load correlation characteristics, the ratio sequence is subtracted point by point from adjacent values ​​to form a difference sequence, and the continuously increasing segments of the difference sequence are extracted and their lengths are calculated to generate interval identification results; Based on the interval identification results, when the interval length exceeds the threshold, the degenerate starting point is recorded at the starting position, and the data segment from the degenerate starting point to the ending position is input into the long short-term memory network for time series modeling. During the modeling process, the offset is accumulated point by point and the sequence features are output to generate the offset accumulation result. Based on the cumulative offset result, the offset sequence is compared with the threshold point by point and the continuous intervals exceeding the threshold are recorded. The intervals exceeding the limit are marked as abnormal paths and the location range is output to generate the degradation trend determination result.

[0010] As a further aspect of the present invention, in the Long Short-Term Memory (LSTM) network, the data segment is input to the input layer of the LSM network. The input layer arranges the values ​​of each point in the ratio difference sequence in chronological order and sends them to the memory unit point by point. The memory unit includes an input gate, a forget gate, and an output gate. The input gate controls whether the current input is written into the unit state, the forget gate controls whether the unit state at the previous time step is retained, and the output gate controls whether the current unit state is output to the next layer. At each time step, the input gate calculates the input weight based on the weighted sum of the current input and the hidden state at the previous time step. The forget gate calculates the forget weight in the same way and performs weighted processing on the unit state at the previous time step to generate an updated unit state. The output gate generates a hidden state based on the updated unit state and passes it to the next time step. At the same time, the offset is accumulated and calculated to form an output sequence. After all time steps are processed, the point-by-point offset values ​​recorded in the output sequence are integrated into an offset accumulation result.

[0011] As a further aspect of the present invention, the specific steps for generating the lifetime prediction distribution results are as follows: Based on the degradation path feature set, the heat dissipation unit temperature rise sequence and magnetic element jitter sequence are divided into time periods and assigned to the state set. The state set is recorded into four types of states, and the state division result is generated. Based on the state division results, the changes between adjacent states are counted as the number of transitions, and the number of transitions is divided by the total number of transitions to obtain the proportion value. The proportion value is filled into the matrix cells according to the row and column positions to form the transition probability matrix. The state transition law is defined by the Markov model to generate the state transition matrix result. Based on the state transition matrix results, the matrix is ​​multiplied by power to obtain a convergent distribution. The distribution values ​​of severe degradation and failure states are compared with the threshold. When the distribution value exceeds the limit, the interval is marked as the critical interval, and the lifetime prediction distribution results are obtained.

[0012] As a further aspect of the present invention, the Markov model first arranges the set of multiple states, which are divided into the temperature rise of the heat dissipation unit and the jitter sequence of the magnetic element, in chronological order. The number of changes between adjacent states is recorded and accumulated item by item. After the statistical results are normalized, they are filled into the matrix units to form a transition probability matrix. The sum of the elements in each row of the transition probability matrix is ​​always equal to 1, which represents the probability distribution of transition from any state to all states. The transition probability matrix is ​​set to a single transition law. The matrix is ​​multiplied by powers, and the matrix obtained by each calculation tends to converge to a convergent distribution. The convergent distribution represents the probability that the long-term state remains stable in all states. When the probability values ​​of the severely degraded state and the failure state are compared with the preset threshold point by point and exceed the limit, the time interval is marked as the critical interval of lifespan, and the lifespan prediction distribution result is output.

[0013] As a further aspect of the present invention, the specific steps for generating the degradation path feature set are as follows: Based on the degradation trend determination result, the filter capacitor ripple sequence and the temperature rise sequence are compared point by point to generate a difference sequence. The difference is squared point by point and then accumulated to form a sum. The sum is allocated to the matrix unit to generate the difference calculation result. Based on the difference calculation results, the matrix units are compared and filtered out of the threshold units one by one. The filtered units are combined and recorded as a set of feature items. The set is reconstructed into an output matrix according to the row and column positions to obtain the degradation path feature set.

[0014] As a further aspect of the present invention, the differences are squared point by point and then accumulated to form a sum. Each difference value is first multiplied by itself to obtain a square value. The square values ​​are recorded point by point in chronological order and accumulated during the recording process. The accumulation operation starts from the beginning of the sequence and ends at the end of the sequence. The accumulated results of all square values ​​are summarized into a total value.

[0015] As a further aspect of the present invention, the specific steps for generating the health assessment conclusion are as follows: Based on the lifetime prediction distribution results, abnormal paths of the control loop are superimposed on the critical interval and the points are marked. The abnormal points are compared with the interval points one by one. The number of matching points and the total number of points are counted to generate the overlap ratio result. Based on the overlap ratio result, the overlap ratio is compared with the threshold successively. When the ratio is greater than the threshold, a maintenance trigger mark is generated. The comparison result and the trigger mark are integrated into the output item to obtain the health assessment conclusion.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the operating load correlation feature is established by calculating the mean and range. In the ratio sequence, the continuously increasing segments are extracted by the difference method. When the interval length exceeds the threshold, the degradation start point is recorded and the data segment is input into the long short-term memory network for time series modeling. During the modeling process, the offset is accumulated point by point and compared with the threshold step by step. The abnormal path range is output and the degradation trend judgment result is generated to ensure that the degradation trend is dynamically tracked. In this invention, the temperature rise of the heat dissipation unit and the vibration sequence of the magnetic element are divided into four states. The changes of adjacent states are statistically analyzed one by one and a proportional matrix is ​​generated. The convergent distribution is obtained by the Markov model through power product. The distribution values ​​of severe degradation and failure states are compared with the threshold point by point. When the threshold is exceeded, the critical interval is marked to form the lifetime prediction distribution, so that the lifetime prediction can combine multi-dimensional state changes rather than rely on a single index. Based on the degradation trend, the filter capacitor ripple and the temperature rise sequence are compared point by point to form a difference sequence, and the sum is obtained by squaring and accumulating. The sum is then assigned to matrix units and filtered to extract the overthreshold units to construct a degradation path feature set. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the main steps of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] Example 1 Please see Figure 1 This invention provides a technical solution: a method for assessing the health of a switching power supply based on data analysis, comprising the following steps: S1: Based on voltage fluctuation sequence, current fluctuation sequence, and power device operating characteristics, time alignment is performed and instantaneous power value is calculated. Then, it is divided point by point with the temperature rise sequence to generate a ratio sequence. The mean and range of the ratio sequence are calculated to establish the operating load correlation characteristics. S2: Based on the characteristics of the operating load association, the difference is calculated segment by segment in the ratio sequence and the rising interval is identified. When the interval exceeds the threshold, the degradation start point is marked and input into the long short-term memory network. The cumulative offset of the data segment is extracted and compared with the threshold point by point. The abnormal path is marked for continuous exceeding intervals, and the degradation trend judgment result is generated. S3: Based on the degradation trend determination results, extract the filter capacitor ripple and temperature rise sequence in the abnormal interval and compare them to generate a difference sequence and calculate the sum of squares to form relevant values. Construct a matrix to filter out the threshold terms and obtain the degradation path feature set. S4: Based on the degradation path feature set, the temperature rise of the heat dissipation unit and the vibration of the magnetic component are assigned to the state set and divided into four categories. The transition is recorded and a ratio matrix is ​​generated. The state transition law is defined by the Markov model. The distribution is obtained by power calculation. When the severe degradation and failure exceed the threshold, the interval is marked as critical, and the lifetime prediction distribution result is obtained. S5: Based on the lifespan prediction distribution results, superimpose abnormal paths of the control loop in the critical interval, calculate the overlap ratio and compare it with the threshold. If the ratio exceeds the limit, generate a trigger flag and obtain the health assessment conclusion.

[0020] The operating load correlation characteristics include the mean power ratio, the range of power ratio, and the power fluctuation amplitude. The degradation trend determination results include the degradation start point location, the cumulative offset value, and the abnormal path range. The degradation path feature set includes the ripple threshold point, the temperature rise threshold point, and the matrix anomaly item. The lifetime prediction distribution results include the state distribution probability, the lifetime critical interval, and the remaining lifetime range. The health assessment conclusions include the lifetime critical point, the abnormal overlap ratio, and the maintenance trigger mark.

[0021] The specific steps for generating operational load correlation characteristics are as follows: Based on voltage fluctuation sequence, current fluctuation sequence, and power device operating characteristics, voltage and current are multiplied point by point to obtain instantaneous power sequence, and instantaneous power is divided point by temperature rise to form ratio sequence, thus generating ratio sequence result; Based on the ratio sequence results, the mean is obtained by summing the values ​​at each point, and the range is obtained by the difference between the valley value and the peak value. The combination of the mean and the range is recorded as a parameter set to establish the operating load correlation characteristics. Based on voltage fluctuation sequences, current fluctuation sequences, and the operating characteristics of power devices, the voltage and current are multiplied point by point using a matrix dot product algorithm. The command is `numpy.multiply`, and the parameters are the voltage sequence array and the current sequence array. The array indices are aligned point by point, and the numerical precision parameter is set to `float64`. During the calculation, each pair of voltage and current data is multiplied item by item to generate an instantaneous power sequence. The instantaneous power sequence is then divided point by point with the temperature rise sequence using a vector division method. The command is `numpy.divide`, and the parameters are the instantaneous power array and the temperature rise array. The array lengths must be the same, and the division method is set to divide corresponding elements point by point. The outlier handling parameter is `ignore`. The division is performed point by point during the calculation, and the results are recorded to generate a ratio sequence. Based on the ratio sequence results, the values ​​of each point in the sequence are summed using the cumulative summation method. The command is `numpy.sum`, the parameter is the ratio sequence array, the precision parameter is `float64`, and the calculation method is to accumulate the values ​​point by point throughout the entire sequence to generate the mean. The mean is calculated using the mean function method, the command is `numpy.mean`, the parameter is the ratio sequence array, and the dimension parameter is `axis=0`. The mean result is generated. The valley and peak values ​​in the ratio sequence are differencing the range using the range calculation method, the command is `numpy.ptp`, the parameter is the ratio sequence array, the dimension parameter is `axis=0`, and the calculation method is to subtract the minimum value from the maximum value to generate the range result. The mean result and the range result are combined and recorded as a parameter set to generate the operational load correlation feature.

[0022] The specific steps for generating degradation trend determination results are as follows: Based on the characteristics of operational load correlation, the ratio sequence is subtracted point by point from adjacent values ​​to form a difference sequence, and the continuously increasing segments of the difference sequence are extracted and their lengths are calculated to generate interval identification results; Based on the interval recognition results, when the interval length exceeds the threshold, the degenerate starting point is recorded at the starting position, and the data segment from the degenerate starting point to the ending position is input into the long short-term memory network for time series modeling. During the modeling process, the offset is accumulated point by point and the sequence features are output to generate the offset accumulation result. Based on the offset accumulation results, the offset sequence is compared with the threshold point by point and the continuous intervals exceeding the threshold are recorded. The intervals exceeding the limit are marked as abnormal paths and the location range is output to generate the degradation trend judgment result. Based on the operational load correlation features, the vector difference algorithm is used to subtract adjacent values ​​in the ratio sequence point by point. The command is numpy.diff, the parameter is the ratio sequence array, the dimension parameter is axis=0, and the numeric type parameter is float64. The operation of subtracting the previous index value from the current index value is performed point by point to generate the difference sequence. The incremental interval extraction method is used to traverse the difference sequence. The command is a for loop structure, the parameter is the sequence index range from 0 to the sequence length minus 1, and the conditional statement is to record the starting position and accumulate the length count when the difference value is greater than 0. Continuous segments are extracted and length statistics are performed to generate interval recognition results. Based on the interval recognition results, a threshold comparison operation is performed on the length value. The threshold determination method is an if statement with parameters of interval length and threshold value. The comparison method is the greater than operator. When the determination is true, the starting index of the interval is recorded as the degeneration start point. The data segment from the degeneration start point to the termination position is input into the Long Short-Term Memory (LSTM) network for temporal modeling. The LSM network input layer command is keras.layers.LSTM with parameters that the input dimension is equal to the data segment length. The number of hidden layer units is 128, the activation function is tanh, the recursive activation function is sigmoid, the time step parameter is point-by-point input, the batch size parameter is 1, the forget gate input weight matrix dimension is 128×128, the input gate and output gate weight matrices are both 128×128, and the bias parameter is initialized to a zero matrix. During the execution process, the offset is accumulated at each time step and the output gate generates the hidden state. The output hidden state sequence is used as the sequence feature to generate the offset accumulation result. Based on the cumulative offset results, a threshold comparison operation is performed point by point on the offset sequence. A point-by-point threshold detection method is adopted, with the command item being numpy.where, the parameters being the offset array and the threshold array, the comparison method being the greater than operator, and the output being an index set. Continuous index segments are extracted from the index set and recorded using an interval marking method. The command item is a for loop and a marker variable, the parameters being the start and end positions of the continuous index, the marker value is set as an abnormal path identifier, and the output is the position range, generating a degradation trend determination result.

[0023] In a Long Short-Term Memory (LSTM) network, data segments are input to the input layer. The input layer arranges the values ​​of each point in the ratio difference sequence in chronological order and sends them to the memory unit point by point. The memory unit includes an input gate, a forget gate, and an output gate. The input gate controls whether the current input is written into the unit state, the forget gate controls whether the unit state of the previous time step is retained, and the output gate controls whether the current unit state is output to the next layer. At each time step, the input gate calculates the input weight based on the weighted sum of the current input and the hidden state of the previous time step. The forget gate calculates the forget weight in the same way and weights the unit state of the previous time step to generate the updated unit state. The output gate generates the hidden state based on the updated unit state and passes it to the next time step. At the same time, the offset is accumulated and calculated to form the output sequence. After all time steps are processed, the point-by-point offset values ​​recorded in the output sequence are integrated into the offset accumulation result. Long Short-Term Memory (LSTM) networks, according to the formula: ; in: Indicates the end position The cumulative result of the switching power supply health deviation at that time. Indicates the starting index of degradation. Indicates the terminating position index. This indicates that the Long Short-Term Memory network is at time 1000. The hidden state vector, Indicates the output projection matrix. This indicates the output layer bias. Indicates time Confidence weight This represents the time reduction factor. Indicates the scaling calibration factor. Indicates the confidence coefficient for the starting point of degradation; Execution process: First, when the interval length exceeds the threshold, the degradation starting point index is determined and denoted as... Then, when the running data meets the termination condition, the termination position index is determined and denoted as... , will the interval The running data segment is input into the Long Short-Term Memory network and at each time step Obtain the hidden state vector Then, the output projection matrix is ​​used. Offset with output layer right A linear transformation is performed to obtain the point-by-point offset, and then confidence weights are introduced into the offset. This is used to suppress the effects of instability, while multiplying by a time reduction factor. This is used to reduce early noise contribution, and then the scaling coefficients are used for calibration. Correcting the overall cumulative bias, combined with the confidence coefficient of the degradation initiation point. To control the reliability of degradation initiation detection, and finally in the interval The weighted offsets at all times are accumulated and then multiplied by 1 in turn. and , obtain the termination position Corresponding cumulative results of switching power supply health offset It is used to reflect the degradation state of the switching power supply within the range and serves as a quantitative indicator for health assessment.

[0024] The specific steps for generating lifetime prediction distribution results are as follows: Based on the degradation path feature set, the heat dissipation unit temperature rise sequence and magnetic element jitter sequence are divided into time periods and assigned to the state set. The state set is recorded into four types of states, generating the state division result. Based on the state partitioning results, the changes between adjacent states are counted as the number of transitions, and the number of transitions is divided by the total number of transitions to obtain the proportion value. The proportion value is filled into the matrix cells according to the row and column positions to form the transition probability matrix. The state transition law is defined by the Markov model to generate the state transition matrix result. Based on the state transition matrix results, the matrix is ​​multiplied by power to obtain a convergent distribution. The distribution values ​​of severe degradation and failure states are compared with the threshold. When the distribution value exceeds the limit, the interval is marked as the critical interval, and the lifetime prediction distribution results are obtained. Based on the degradation path feature set, the temperature rise sequence of the heat dissipation unit and the vibration sequence of the magnetic component are divided into time periods. The interval division method is adopted, with the command item being numpy.array_split, the parameters being the input sequence array and the interval value, the interval value being set to 4, the division method being equal length splitting, and the output being multiple subsequences. The state assignment method is used to assign values ​​to each subsequence, with the command item being a for loop statement, the parameters being the subsequence index range 0 to 3, and the assignment rule being that index 0 is assigned to state one, index 1 is assigned to state two, index 2 is assigned to state three, and index 3 is assigned to state four. The assignment results are recorded sequentially as a state set to generate the state division result. Based on the state partitioning results, statistical operations are performed on the changes between adjacent states. A transition count method is used, with the command `collections.Counter`. The input parameter is the combination of two adjacent points in the state sequence. The statistical method involves traversing the entire sequence and recording the number of combinations for each pair of adjacent states. The output is a dictionary-style set of transition counts. A proportional calculation method is used to divide the transition counts by the total number of times. The command `numpy.divide` takes the transition count array and the total number of times as parameters. The division method is item-by-item division, generating a proportional value array. A matrix filling method is used to fill the proportional value array into matrix cells, with the command `numpy.reshape` taking the proportional value array and a matrix shape of four times four as parameters. The filling order parameter is row-major. The output is the transition probability matrix, which is then input into the Markov model method. The command `numpy.matmul` performs matrix multiplication, taking the transition probability matrix and the current state vector as parameters, generating the state transition matrix result. Based on the state transition matrix results, a power-law accumulation operation is performed on the matrix using the matrix power calculation method. The command is `numpy.linalg.matrix_power`, with parameters being the state transition matrix and the power value, which is set to 100. The operation involves successive multiplication of the matrix with itself, and the output is a convergent matrix distribution. A threshold comparison method is then used to compare the probability values ​​of severely degraded and failed states in the convergent distribution point by point. The command is `numpy.greater`, with parameters being the distribution array and the threshold array. The comparison method is a successive greater than operator, and the output is a set of Boolean indices. The intervals corresponding to the Boolean index set are marked as critical intervals, generating the lifetime prediction distribution results.

[0025] The Markov model first arranges the set of multiple states, which are divided into the temperature rise of the heat dissipation unit and the vibration sequence of the magnetic element, in chronological order. The number of changes between adjacent states is recorded and accumulated item by item. After the statistical results are normalized, they are filled into the matrix cells to form the transition probability matrix. The sum of the elements in each row of the transition probability matrix is ​​always equal to 1, which represents the probability distribution of transition from any state to all states. The transition probability matrix is ​​set to a single transition law. The matrix is ​​multiplied by power, and the matrix obtained by each calculation tends to converge to a convergent distribution. The convergent distribution represents the probability that the long-term state remains stable in all states. When the probability values ​​of the severely degraded state and the failure state are compared with the preset threshold point by point and exceed the limit, the time interval is marked as the lifetime critical interval, and the lifetime prediction distribution result is output. Markov models, according to the formula: ; in: This indicates the status during the health assessment process. Transition to state Improved transition probability, Indicates at time The operating status of the switching power supply This represents the total length of the state sequence. Represents the set of all possible running states. Indicates the initial state of the transition. Indicates the target state of the transition. This represents the candidate target states traversed in the normalized denominator. Indicates when the state is at time 10:00 for And at the moment for An indicator function that takes a value of 1 if the condition is met and a value of zero otherwise. Indicates time The time weighting is used to enhance the contribution of recent state transitions to health assessment. Indicates from state to state The stability coefficient is used to suppress low-confidence transitions. The Dirichlet smoothing intensity is used to avoid zero-probability cases. Indicates the target state The prior probability components are used to provide a smoothing benchmark for transition probabilities; Execution process: First, collect a running state sequence of length L and record the state at each time step as . Then determine the set of states. And select the start state of the transfer. With the target state of the transfer Then, traversing the time index to Using indicator functions Determine if there exists a state to state The transition is performed and the result is returned as zero or one, then multiplied by the time weight at each valid transition. Increase the contribution of recent transfers and multiply by the stability coefficient. To weaken low-confidence transitions, the weighted count results are then summed at the numerator and a smoothing term is added. For all candidate target states on the denominator side Perform the same calculation to obtain the weighted normalization factor, and finally divide the numerator by the denominator to calculate the transition probability. and all The calculation results are filled into the state transition matrix according to the row and column positions, which serves as the key result describing the state evolution law in the health assessment of switching power supplies.

[0026] The specific steps for generating the feature set of the degradation path are as follows: Based on the degradation trend determination result, the filter capacitor ripple sequence and the temperature rise sequence are compared point by point to generate a difference sequence. The difference is squared point by point and then accumulated to form a sum. The sum is distributed to the matrix unit to generate the difference calculation result. Based on the difference calculation results, the matrix unit is compared and filtered out the over-threshold unit one by one. The filtered units are combined and recorded as a feature term set. The set is reconstructed into an output matrix according to the row and column positions to obtain the degradation path feature set. Based on the degradation trend determination results, the filter capacitor ripple sequence and the temperature rise sequence are compared point by point using a point-by-point difference calculation method. The command is numpy.subtract, and the parameters are the filter capacitor ripple array and the temperature rise array. The array lengths must be consistent, and the index positions must correspond point by point. The calculation method is to subtract the temperature rise value from the ripple value at the current index point. The output is a difference sequence. The difference sequence is then calculated using a point-by-point square accumulation method. The commands are numpy.power and numpy.cumsum, and the parameters are the difference sequence array and the power value 2. The power operation is performed point by point, and the squared results are input into the accumulation command in chronological order. The accumulation method is point-by-point accumulation, with the starting point being the first element of the sequence and the ending point being the last element of the sequence. The accumulation result is recorded as a sum. The sum is then allocated using a matrix cell allocation method. The command is numpy.reshape, and the parameters are the sum array and the matrix shape m×n. The filling method is row-major, generating the difference calculation result. Based on the difference calculation results, comparison operations are performed on each matrix element using a threshold filtering method. The command is `numpy.greater`, with parameters being the matrix element array and the threshold array. The comparison method is a pointwise greater than operator, and the output is a Boolean matrix index. The filtering element combination method is used to extract elements with true Boolean indices. The command is `numpy.where`, with parameters being Boolean matrix indices. The output is a set of row and column positions of elements that meet the conditions, which is recorded as a feature set. The feature set is then subjected to a matrix reconstruction method using the command `numpy.zeros` and a for loop, with parameters being the number of matrix rows and columns. The extracted feature items are filled into the corresponding positions, and the output is the reconstructed matrix, resulting in the degenerate path feature set.

[0027] The differences are squared point by point and then summed to form a total. Each difference value is multiplied by itself to obtain the square value. The square values ​​are recorded point by point in chronological order and accumulated during the recording process. The accumulation operation starts from the beginning of the sequence and ends at the end of the sequence. The accumulated results of all square values ​​are summarized into a total value.

[0028] The specific steps for generating a health assessment conclusion are as follows: Based on the lifetime prediction distribution results, the abnormal paths of the control loop are superimposed on the critical interval and the points are marked. The abnormal points are compared with the points in the interval one by one. The number of matching points and the total number of points are counted and the overlap ratio results are generated. Based on the overlap ratio result, the overlap ratio is compared with the threshold successively. When the ratio is greater than the threshold, a maintenance trigger flag is generated. The comparison result and the trigger flag are integrated into the output item to obtain the health assessment conclusion. Based on the lifetime prediction distribution results, an overlay operation is performed on the critical interval and the abnormal path of the control loop. An array-based logical overlay method is used, with the command `numpy.logical_or` and parameters being a Boolean array of the critical interval and an abnormal path. A logical summation operation is performed point-by-point to generate the overlay Boolean array. A point-marking method is used to traverse the overlay Boolean array, with the command `for` loop and parameters ranging from Boolean array index 0 to array length minus 1. A conditional statement records the index position as a point set when the Boolean value is true, generating an abnormal point set. The abnormal point set and the interval point set are then compared. The comparison is performed item by item using the set intersection method. The command is `set.intersection`, the parameters are the set of outlier points and the set of interval points, and the output is the set of matching points. The counting method is used to perform counting operations on the set of matching points and the total set of points respectively. The command is `len`, the parameter is a set object, and the output is the number of matching points and the total number of points. These two are used as input to perform the ratio calculation method. The command is `numpy.divide`, the parameters are the number of matching points and the total number of points, the calculation method is floating-point division, and the result is the overlap ratio. The overlap ratio result is generated. Based on the overlap ratio result, a comparison operation is performed successively between the overlap ratio and the threshold. The successive threshold comparison method is adopted, with the command item being a for loop and an if statement, and the parameters being the overlap ratio value and the threshold value. The comparison method is the greater than operator. When the comparison result is true, the marker generation method is called, with the command item being the append function, and the parameter being the maintenance trigger marker string. The operation is to add the maintenance trigger marker to the output sequence. The comparison result and the trigger marker are merged using the result integration method, with the command item being the dictionary constructor dict, and the parameter key-value pairs including "comparison result" and "trigger marker". The output item is generated to obtain the health assessment conclusion.

[0029] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A data analysis-based method for assessing the health of a switching power supply, characterized in that, Includes the following steps: S1: Based on voltage fluctuation sequence, current fluctuation sequence, and power device operating characteristics, time alignment is performed and instantaneous power value is calculated. Then, it is divided point by point with the temperature rise sequence to generate a ratio sequence. The mean and range of the ratio sequence are calculated to establish the operating load correlation characteristics. S2: Based on the operating load correlation characteristics, calculate the difference segment by segment in the ratio sequence and identify the rising interval. When the interval exceeds the threshold, mark the degradation start point and input it into the long short-term memory network. Extract the cumulative offset of the data segment and compare it with the threshold point by point. Mark the abnormal path in the continuous exceeding interval and generate the degradation trend judgment result. S3: Based on the degradation trend determination result, extract the filter capacitor ripple and temperature rise sequence in the abnormal interval and compare them to generate a difference sequence and calculate the sum of squares to form relevant values. Construct a matrix to filter out the threshold terms and obtain the degradation path feature set. S4: Based on the set of degradation path features, the temperature rise of the heat dissipation unit and the jitter of the magnetic element are assigned to the state set and divided into four categories. The transition is recorded and a ratio matrix is ​​generated. The state transition law is defined by the Markov model. The distribution is obtained by power calculation. When the severe degradation and failure exceed the threshold, the interval is marked as critical to obtain the lifetime prediction distribution result. S5: Based on the lifespan prediction distribution results, superimpose abnormal paths of the control loop in the critical interval, calculate the overlap ratio and compare it with the threshold. If the ratio exceeds the limit, generate a trigger flag and obtain the health assessment conclusion.

2. The data analysis-based health assessment method for switching power supplies according to claim 1, characterized in that, The operating load correlation features include the power ratio mean, power ratio range, and power fluctuation amplitude; the degradation trend determination results include the degradation start point location, cumulative offset value, and abnormal path range; the degradation path feature set includes ripple threshold point, temperature rise threshold point, and matrix anomaly item; the lifetime prediction distribution results include state distribution probability, lifetime critical interval, and remaining lifetime range; and the health assessment conclusions include lifetime critical point, abnormal overlap ratio, and maintenance trigger marker.

3. The data analysis-based health assessment method for switching power supplies according to claim 1, characterized in that, The specific steps for generating the aforementioned operational load correlation features are as follows: Based on voltage fluctuation sequence, current fluctuation sequence, and power device operating characteristics, voltage and current are multiplied point by point to obtain instantaneous power sequence, and instantaneous power is divided point by temperature rise to form ratio sequence, thus generating ratio sequence result; Based on the ratio sequence results, the mean is obtained by summing the values ​​at each point, and the range is obtained by differentiating the valley and peak values. The combination of the mean and range is recorded as a parameter set to establish the operating load correlation characteristics.

4. The data analysis-based health assessment method for switching power supplies according to claim 1, characterized in that, The specific steps for generating the degradation trend determination result are as follows: Based on the aforementioned operating load correlation characteristics, the ratio sequence is subtracted point by point from adjacent values ​​to form a difference sequence, and the continuously increasing segments of the difference sequence are extracted and their lengths are calculated to generate interval identification results; Based on the interval identification results, when the interval length exceeds the threshold, the degenerate starting point is recorded at the starting position, and the data segment from the degenerate starting point to the ending position is input into the long short-term memory network for time series modeling. During the modeling process, the offset is accumulated point by point and the sequence features are output to generate the offset accumulation result. Based on the cumulative offset result, the offset sequence is compared with the threshold point by point and the continuous intervals exceeding the threshold are recorded. The intervals exceeding the limit are marked as abnormal paths and the location range is output to generate the degradation trend determination result.

5. The data analysis-based health assessment method for switching power supplies according to claim 4, characterized in that, In the Long Short-Term Memory (LSTM) network, the data segment is input to the input layer. The input layer arranges the values ​​of each point in the ratio difference sequence in chronological order and sends them to the memory unit point by point. The memory unit includes an input gate, a forget gate, and an output gate. The input gate controls whether the current input is written into the unit state, the forget gate controls whether the unit state at the previous time step is retained, and the output gate controls whether the current unit state is output to the next layer. At each time step, the input gate calculates the input weight based on the weighted sum of the current input and the hidden state at the previous time step. The forget gate calculates the forget weight in the same way and weights the unit state at the previous time step to generate the updated unit state. The output gate generates the hidden state based on the updated unit state and passes it to the next time step. At the same time, the offset is accumulated and calculated to form the output sequence. After all time steps are processed, the point-by-point offset values ​​recorded in the output sequence are integrated into the offset accumulation result.

6. The data analysis-based health assessment method for switching power supplies according to claim 1, characterized in that, The specific steps for generating the lifetime prediction distribution results are as follows: Based on the degradation path feature set, the heat dissipation unit temperature rise sequence and magnetic element jitter sequence are divided into time periods and assigned to the state set. The state set is recorded into four types of states, and the state division result is generated. Based on the state division results, the changes between adjacent states are counted as the number of transitions, and the number of transitions is divided by the total number of transitions to obtain the proportion value. The proportion value is filled into the matrix cells according to the row and column positions to form the transition probability matrix. The state transition law is defined by the Markov model to generate the state transition matrix result. Based on the state transition matrix results, the matrix is ​​multiplied by power to obtain a convergent distribution. The distribution values ​​of severe degradation and failure states are compared with the threshold. When the distribution value exceeds the limit, the interval is marked as the critical interval, and the lifetime prediction distribution results are obtained.

7. The data analysis-based health assessment method for switching power supplies according to claim 6, characterized in that, The Markov model first arranges the set of multiple states, which are divided into the temperature rise of the heat dissipation unit and the jitter sequence of the magnetic element, in chronological order. The number of changes between adjacent states is recorded and accumulated item by item. After the statistical results are normalized, they are filled into the matrix cells to form a transition probability matrix. The sum of the elements in each row of the transition probability matrix is ​​always equal to 1, which represents the probability distribution of transition from any state to all states. The transition probability matrix is ​​set to a single transition law. The matrix is ​​multiplied by powers, and the matrix obtained by each calculation tends to converge to a convergent distribution. The convergent distribution represents the probability that the long-term state remains stable in all states. When the probability values ​​of the severely degraded state and the failure state are compared with the preset threshold point by point and exceed the limit, the time interval is marked as the critical interval of lifespan, and the lifespan prediction distribution result is output.

8. The data analysis-based health assessment method for switching power supplies according to claim 1, characterized in that, The specific steps for generating the degradation path feature set are as follows: Based on the degradation trend determination result, the filter capacitor ripple sequence and the temperature rise sequence are compared point by point to generate a difference sequence. The difference is squared point by point and then accumulated to form a sum. The sum is allocated to the matrix unit to generate the difference calculation result. Based on the difference calculation results, the matrix units are compared and filtered out of the threshold units one by one. The filtered units are combined and recorded as a set of feature items. The set is reconstructed into an output matrix according to the row and column positions to obtain the degradation path feature set.

9. The data analysis-based health assessment method for switching power supplies according to claim 8, characterized in that, The differences are squared point by point and then summed to form a total. Each difference value is multiplied by itself to obtain a square value. The square values ​​are recorded point by point in chronological order and accumulated during the recording process. The accumulation operation starts from the beginning of the sequence and ends at the end of the sequence. The accumulated results of all square values ​​are summarized into a total value.

10. The data analysis-based health assessment method for switching power supplies according to claim 1, characterized in that, The specific steps for generating the health assessment conclusion are as follows: Based on the lifetime prediction distribution results, abnormal paths of the control loop are superimposed on the critical interval and the points are marked. The abnormal points are compared with the interval points one by one. The number of matching points and the total number of points are counted to generate the overlap ratio result. Based on the overlap ratio result, the overlap ratio is compared with the threshold successively. When the ratio is greater than the threshold, a maintenance trigger mark is generated. The comparison result and the trigger mark are integrated into the output item to obtain the health assessment conclusion.

Images