Intelligent operation monitoring method and device of PEM fuel cell

Abstract

The invention discloses an intelligent operation monitoring method and device for a PEM fuel cell, and relates to the technical field of intelligent monitoring. The method comprises the following steps: preprocessing voltage data and temperature data to eliminate dimensional difference and numerical range interference between original data, and then adopting a double-path analysis strategy for the preprocessed data: 1, obtaining a first fault rate through data fault analysis; and secondly, the second fault rate is obtained by substituting the image coding, flattening and reconstructing operation into the light-weight long-short-term network model, and finally the operation state monitoring result is obtained according to the first fault rate and the second fault rate, so that the operation state is evaluated from multiple dimensions, and the accuracy of operation monitoring is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent monitoring, and particularly relates to an intelligent operation monitoring method and device for a PEM fuel cell. BACKGROUND

[0002] PEM fuel cells play a key role in new energy vehicles, distributed energy stations and other fields due to their high energy conversion efficiency, fast response and zero emission characteristics. As the application scenarios of PEM fuel cells continue to expand, accurate monitoring of the operating state of PEM fuel cells has become a core requirement to ensure the stability of energy supply and the safety of equipment. During operation, voltage fluctuations, sudden temperature changes and other faults not only reduce the performance of the battery, but also may cause safety accidents. The existing PEM fuel cell operation monitoring technology has significant defects in accuracy. The traditional threshold judgment method only relies on a single voltage or temperature data compared with a fixed threshold, and cannot consider the coupling relationship between the data and the influence of the working condition change.

[0003] Patent CN116995280A discloses an intelligent monitoring method for the operating state of a hydrogen fuel cell, and relates to the field of battery monitoring. The method includes obtaining a predetermined monitoring variable set of a first hydrogen fuel cell, the predetermined monitoring variable set including M predetermined monitoring variables; building M battery state variable sensing monitoring modules; obtaining M variable monitoring data sets based on the M battery state variable sensing monitoring modules; performing deviation analysis on the M variable monitoring data sets based on a variable state deviation analysis function to obtain M variable monitoring deviation analysis results; detecting the M variable monitoring deviation analysis results with a battery state detection module to obtain a battery state detection result; and obtaining a battery state warning signal when the battery state detection result does not meet a predetermined battery state constraint. The method solves the technical problems of low accuracy and low efficiency of state monitoring and warning of hydrogen fuel cells in the prior art. However, in actual application, the coupling relationship between voltage and temperature data is not fully considered, and the influence of working condition change is not adapted, so it is difficult to further improve the accuracy of operation monitoring in complex operating scenarios. SUMMARY

[0004] The purpose of the present application is to solve the problem of low accuracy of operating state monitoring, and to provide an intelligent operation monitoring method and device for a PEM fuel cell.

[0005] In the first aspect of the present application, an intelligent operation monitoring method for a PEM fuel cell is first proposed, which comprises: synchronously collecting voltage data and temperature data of the PEM fuel cell; respectively pre-processing the voltage data and the temperature data to obtain effective voltage data and effective temperature data; performing data fault analysis on the effective voltage data and the effective temperature data to obtain a first failure rate; performing two-dimensional data conversion on the effective voltage data and the effective temperature data by an image coding operation to obtain a two-dimensional image; performing a flattened dimension reconstruction operation on the two-dimensional image to obtain a time sequence characteristic sequence, and substituting the time sequence characteristic sequence into a light long short-term network model to obtain a second failure rate; obtaining a running state monitoring result according to the first failure rate and the second failure rate.

[0006] Optionally, performing normalization on the voltage data and the temperature data to obtain normalized voltage data and normalized temperature data; dividing the normalized voltage data and the normalized temperature data by a preset sliding window to obtain a sub-voltage data set and a sub-temperature data set; replacing average values of sub-voltage data and sub-temperature data in each window with data points in the window to obtain effective voltage data and effective temperature data.

[0007] Optionally, dividing the effective voltage data and the effective temperature data by a preset sliding window to obtain a sub-effective voltage data set and a sub-effective temperature data set; calculating voltage slopes and temperature slopes of sub-effective voltage data and sub-effective temperature data in each window, and if an absolute value of the voltage slope is greater than a preset voltage slope threshold or an absolute value of the temperature slope is greater than a preset temperature slope threshold, marking data points in the window and obtaining a number of failure points; taking a ratio of all failure point numbers to data point numbers as the first failure rate.

[0008] Optionally, arranging the effective voltage data and the effective temperature data in a time stamp order to obtain a one-dimensional sequence set; performing normalization on one-dimensional data in the one-dimensional sequence set to obtain a normalized data set; performing a preset inverse cosine normalization operation on normalized data in the normalized data set to obtain a two-dimensional feature matrix; performing a preset pixel gray color value mapping on values in the two-dimensional feature matrix to obtain a two-dimensional image.

[0009] Optionally, performing a compressed matrix operation on a weight matrix of a long short-term memory network to obtain a low-rank approximate matrix; the low-rank approximate matrix replaces an original weight matrix of the long short-term network to participate in calculation to obtain a light long short-term network; performing light long short-term network processing on the time sequence characteristic sequence from a start end of the time sequence characteristic sequence forward and from an end of the time sequence characteristic sequence backward in each time step to obtain a forward state vector and a backward state vector. The forward state vector and the reverse state vector are spliced to obtain a fusion state vector; The fusion feature vector is nonlinearly transformed by a preset monitoring formula to obtain a second failure rate.

[0010] In the second aspect of the embodiment of the present application, an intelligent operation monitoring device for a PEM fuel cell is provided, comprising: A data acquisition module is configured to acquire voltage data and temperature data of the PEM fuel cell; the voltage data and the temperature data are data acquired synchronously in a preset period before the current time; A preprocessing module is configured to preprocess the voltage data and the temperature data to obtain effective voltage data and effective temperature data; A first failure rate generation module is configured to perform data failure analysis on the effective voltage data and the effective temperature data to obtain a first failure rate; A two-dimensional image generation module is configured to perform two-dimensional data conversion on the effective voltage data and the effective temperature data by image encoding operation to obtain a two-dimensional image; A second failure rate generation module is configured to perform flattening dimension reconstruction operation on the two-dimensional image to obtain a time sequence feature sequence, and to obtain a second failure rate by substituting the time sequence feature sequence into a lightweight long short-term network model; A detection result generation module is configured to obtain an operation state monitoring result according to the first failure rate and the second failure rate.

[0011] Optionally, the preprocessing module comprises: A first normalization module is configured to normalize the voltage data and the temperature data to obtain normalized voltage data and normalized temperature data; A first data segmentation module is configured to divide the normalized voltage data and the normalized temperature data by a first preset sliding window to obtain a sub-voltage data set and a sub-temperature data set; A data replacement module is configured to replace the average value of the sub-voltage data and the sub-temperature data in each window with the data points in the window to obtain effective voltage data and effective temperature data.

[0012] Optionally, the first failure rate generation module comprises: A second data segmentation module is configured to divide the effective voltage data and the effective temperature data by a second preset sliding window to obtain a sub-effective voltage data set and a sub-effective temperature data set; a marking module, configured to calculate a voltage slope and a temperature slope of the sub-effective voltage data and the sub-effective temperature data in each window, and mark data points in the window and obtain a number of fault points if an absolute value of the voltage slope is greater than a preset voltage slope threshold or an absolute value of the temperature slope is greater than a preset temperature slope threshold; a first failure rate module, configured to take a ratio of the number of fault points to the number of data points as a first failure rate.

[0013] Optionally, the two-dimensional image generation module comprises: a sorting module, configured to arrange the effective voltage data and the effective temperature data in a time stamp order to obtain a one-dimensional sequence set; a second normalization module, configured to normalize one-dimensional data in the one-dimensional sequence set to obtain a normalized data set; a third normalization module, configured to perform a preset arccosine normalization operation on the normalized data in the normalized data set to obtain a two-dimensional feature matrix; a mapping module, configured to perform a preset pixel gray color value mapping on values in the two-dimensional feature matrix to obtain a two-dimensional image.

[0014] Optionally, the second failure rate generation module comprises: a compression module, configured to perform a compression matrix operation on a weight matrix of the long short-term memory network to obtain a low-rank approximation matrix; a replacement module, configured to replace the low-rank approximation matrix with an original weight matrix of the long short-term network to participate in calculation to obtain a light long short-term network; a bidirectional processing module, configured to perform light long short-term network processing on the time sequence feature sequence from a start end of the time sequence feature sequence in a forward direction and from an end of the time sequence feature sequence in a reverse direction step by step to obtain a forward state vector and a reverse state vector; a splicing module, configured to splice the forward state vector and the reverse state vector to obtain a fusion state vector; a second failure rate module, configured to perform a nonlinear transformation on the fusion feature vector through a preset monitoring formula to obtain a second failure rate.

[0015] The beneficial effects of the present application are: synchronously collecting voltage data and temperature data of a PEM fuel cell; respectively pre-processing the voltage data and the temperature data to obtain effective voltage data and effective temperature data; performing data fault analysis on the effective voltage data and the effective temperature data to obtain a first fault rate; performing two-dimensional data conversion on the effective voltage data and the effective temperature data through image coding to obtain a two-dimensional image; performing a flattened dimension reconstruction operation on the two-dimensional image to obtain a time sequence feature sequence, and substituting the time sequence feature sequence into a lightweight long short-term network model to obtain a second fault rate; and obtaining a running state monitoring result according to the first fault rate and the second fault rate. Through pre-processing of the voltage data and the temperature data, the dimensional difference and the numerical range interference between original data are eliminated, and then a double-path analysis strategy is used on the pre-processed data: one is to obtain the first fault rate through data fault analysis; the other is to substitute the two-dimensional image obtained through image coding and flattened reconstruction into the lightweight long short-term network model to obtain the second fault rate, and finally the running state monitoring result is obtained according to the first fault rate and the second fault rate, so that the running state is evaluated from multiple dimensions, and the accuracy of the running monitoring is improved. BRIEF DESCRIPTION OF DRAWINGS

[0016] The present application will be further described below in conjunction with the accompanying drawings.

[0017] Figure 1 A flowchart of an intelligent running monitoring method of a PEM fuel cell provided by an embodiment of the present application; Figure 2 A flowchart of another intelligent running monitoring method of a PEM fuel cell provided by an embodiment of the present application; Figure 3 A framework diagram of an intelligent running monitoring device of a PEM fuel cell provided by an embodiment of the present application; DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application.

[0019] Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0020] An intelligent running monitoring method of a PEM fuel cell is provided by an embodiment of the present application. Referring to Figure 1 , Figure 1 A flowchart of an intelligent running monitoring method of a PEM fuel cell provided by an embodiment of the present application. The method comprises the following steps: S101, collecting voltage data and temperature data of a PEM fuel cell; S102, respectively, pre-process the voltage data and the temperature data to obtain effective voltage data and effective temperature data; S103, performing data fault analysis on the effective voltage data and the effective temperature data to obtain a first failure rate; S104, performing two-dimensional data conversion on the effective voltage data and the effective temperature data by image coding operation to obtain a two-dimensional image; S105, performing a flattening dimension reconstruction operation on the two-dimensional image to obtain a time sequence feature sequence, and substituting the time sequence feature sequence into a lightweight long short-term network model to obtain a second failure rate; S106, obtaining a running state monitoring result according to the first failure rate and the second failure rate.

[0021] The intelligent running monitoring method of the PEM fuel cell provided in the embodiment of the application eliminates the dimensional difference and the numerical range interference between the original data by pre-processing the voltage data and the temperature data, and then adopts a double-path analysis strategy on the pre-processed data: one is to obtain a first failure rate by data fault analysis; the other is to obtain a second failure rate by substituting a two-dimensional image obtained by image coding and flattening reconstruction operation into a lightweight long short-term network model, and finally to obtain a running state monitoring result according to the first failure rate and the second failure rate, so as to evaluate the running state from multiple dimensions and improve the accuracy of running monitoring.

[0022] In an implementation manner, the flattening dimension reconstruction operation is a convolution operation on the two-dimensional image with a fixed size, a multi-channel feature map is obtained, a fixed dimension feature map is obtained by maximum pooling on the multi-channel feature map, and a time sequence feature sequence is obtained by sequentially arranging each element of the fixed dimension feature map.

[0023] In an implementation manner, a comprehensive failure rate is calculated according to the first failure rate and the second failure rate, the comprehensive failure rate = α * the first failure rate + (1-α) * the second failure rate, if the comprehensive failure rate >= threshold value A, the running state is abnormal, if the comprehensive failure rate < threshold value A, the running state is normal, wherein the threshold value A is determined by a technician.

[0024] In an implementation manner, the voltage data and the temperature data are synchronously collected, so that the dynamic change trend of the battery in a period of time can be completely captured, and accidental error of single-point data can be avoided.

[0025] In an implementation manner, the original data may contain abnormal values generated by sensor errors or system fluctuations, and the pre-processing can improve the data reliability.

[0026] In one implementation, a preset sliding window is used to calculate voltage and temperature slopes, and a fault point is marked by comparing with a threshold value to quickly lock the abnormal trend of data, the proportion of the fault point is quantified as a first fault rate to provide an intuitive index for rapid evaluation of the operating state, the voltage and temperature data are converted into a two-dimensional image to realize the conversion from time domain to spatial features, the data distribution law is highlighted by virtue of the image visualization advantage, and the feature discrimination is enhanced by combining the inverse cosine normalization and pixel mapping, and the fault recognition capability is improved.

[0027] In one implementation, one-dimensional time series data is converted into a two-dimensional image, the time sequence correlation and spatial features of the data are preserved, and richer input features are provided for a deep learning model. Image coding can compress high-dimensional time series data into a two-dimensional matrix, reduce data redundancy, while preserving key features, improve the calculation efficiency of subsequent models, fuse the fault rates obtained by traditional calculation and deep learning, realize the complementation of rapid preliminary judgment and deep diagnosis, comprehensively improve the accuracy of operating state evaluation, reduce false positives and false negatives, and provide a reliable basis for monitoring results.

[0028] In one embodiment, the voltage data and the temperature data are normalized to obtain normalized voltage data and normalized temperature data; The normalized voltage data and the normalized temperature data are divided by a first preset sliding window to obtain a sub-voltage data set and a sub-temperature data set; The average value of the sub-voltage data and the sub-temperature data in each window is replaced with the data points in the window to obtain effective voltage data and effective temperature data.

[0029] In one implementation, all data points of the voltage data and the temperature data are traversed, the voltage mean μ1 and the temperature mean μ2 are calculated, and then the voltage standard deviation σ1 and the temperature standard deviation σ2 of the voltage data and the temperature data are calculated. For each voltage data point Vi, the normalized value Vi is calculated as follows: For each temperature data point Tj, the normalized value Tj is calculated as follows: .

[0030] In one implementation, the first preset sliding window size n is determined by a technician, the first to the n-th data points in the normalized voltage data are taken to calculate the mean value, and the original values of the first to the n-th data points are replaced; the window slides to the right by 1 data point, the 2nd to the n+1th data points are taken to calculate the mean value, and the values of the 2nd to the n+1th points are replaced; this process is repeated until all data are traversed, and the processing operation on the normalized temperature data is the same.

[0031] In one embodiment, the effective voltage data and the effective temperature data are divided by a second preset sliding window to obtain a sub-effective voltage data set and a sub-effective temperature data set; Calculate the voltage slope and temperature slope of the sub-effective voltage data and sub-effective temperature data in each window, and if the absolute value of the voltage slope is greater than the preset voltage slope threshold or the absolute value of the temperature slope is greater than the preset temperature slope threshold, mark the data points in the window and obtain the number of fault points. The ratio of all fault point numbers to data point numbers is taken as the first failure rate.

[0032] In an implementation, the second preset sliding window m, the preset voltage threshold and the preset temperature threshold are determined by technicians, the step is 1 data point position, the voltage slope of the effective voltage data V1, V2,..., Vm in each window can be fitted with a straight line V=a*t+b, a is the voltage slope, and the time indexes corresponding to the m data points in the window are t1, t2,..., tm. m Then the slope a= The temperature slope of the effective temperature data T1, T2,..., Tm in each window is calculated in the same way.

[0033] In an implementation, the mutation or abnormal accelerated change of the data can be captured in real time by piecewise calculation through the sliding window, and the fixed window size can balance the short-term fluctuation sensitivity and the long-term trend stability, the data points in the window are marked when the absolute value of the slope exceeds the threshold, the time interval when the failure occurs can be directly located, and accurate positioning is realized.

[0034] In one embodiment, referring to Figure 2 On the basis of Figure 1 After step S104, the method further includes: S1041, arranging the effective voltage data and the effective temperature data in time stamp order to obtain a one-dimensional sequence set; S1042, normalizing the one-dimensional data in the one-dimensional sequence set to obtain a normalized data set; S1043, performing a preset inverse cosine normalization operation on the normalized data in the normalized data set to obtain a two-dimensional feature matrix; S1044, performing pixel gray mapping on the values in the two-dimensional feature matrix to obtain a two-dimensional image.

[0035] In an implementation, the sorting in time stamp order means that the collected effective voltage data and effective temperature data are arranged in one-dimensional sequence set in the order of voltage and temperature according to the time sequence before and after each data point.

[0036] In an implementation, the normalization operation can be linear normalization, Z-Score standardization, etc., and the preset inverse cosine normalization operation is obtained according to the historical experience of the staff.

[0037] In one implementation, the matrix numerical range is linearly mapped to the interval of 0-255, and the gray value is calculated by the formula, and the gray value is calculated by the formula .

[0038] In one implementation, the normalized data meets the input requirements of the machine learning model, which can reduce the shock during gradient update, improve the model training efficiency and feature extraction effect, and the inverse cosine normalization can convert the linear relationship of one-dimensional data into a nonlinear distribution in two-dimensional space, exposing the hidden features in the original data. After mapping the numerical value to the gray value, the two-dimensional image can intuitively show the distribution pattern of the data, which not only preserves the time sequence information of the original data, but also mines the hidden correlation features through spatial transformation and image processing.

[0039] In one embodiment, the weight matrix of the long short-term memory network is subjected to a compression matrix operation to obtain a low-rank approximation matrix; The low-rank approximation matrix replaces the original weight matrix of the long short-term network to participate in calculation to obtain a lightweight long short-term network; The time sequence feature sequence is subjected to lightweight long short-term network processing from the start of the time sequence feature sequence and from the end of the time sequence feature sequence in reverse order to obtain a forward state vector and a reverse state vector; The forward state vector and the reverse state vector are spliced to obtain a fusion state vector; The fusion feature vector is subjected to nonlinear transformation by a preset monitoring formula to obtain a second failure rate; The forward state vector and the reverse state vector are spliced to obtain a fusion state vector; The fusion feature vector is subjected to nonlinear transformation by a preset monitoring formula to obtain a second failure rate.

[0040] In one implementation, after inputting the time sequence feature sequence, the process of processing the time sequence feature sequence from the start in a forward direction step by step includes: first, singular value decomposition is performed on the weight matrix involved, which is decomposed into a left singular matrix, a singular value matrix, and a right singular matrix; then, the first k largest singular values are selected (which can be dynamically adjusted by a technician according to actual conditions), and a low-rank approximation matrix is reconstructed using these selected singular values, which is used to participate in calculation to generate a forward state vector; the process of processing the time sequence feature sequence from the end in a reverse direction step by step also performs the above steps, but dynamically improves the singular value retention proportion of the weight corresponding to the end time step to strengthen the capture of key information in the reverse time sequence to generate a reverse state vector. Compared with the lightweight long short-term memory network, in terms of matrix processing mechanism, the traditional long short-term memory network directly participates in calculation using a full-rank matrix, and the dimension of the weight matrix is W , the parameter redundancy is high, the calculation complexity reaches O(mn), and there is no decomposition optimization measure, relies on model compression or quantization and other post-processing means, and the network structure is easily damaged; and the new scheme performs singular value decomposition on the weight matrix of the forward and reverse directions, reconstructs a low-rank matrix by retaining the first k largest singular values, and reduces the calculation complexity to O(mk+kn).

[0041] In an implementation manner, the second failure rate =σ ), wherein σ can be an activation function Sigmoid, ReLU, etc., 2 indicates that there are 2 subspace characteristic voltages and temperatures, represents an important coefficient of the i-th subspace characteristic, is a characteristic weighting coefficient of the i-th subspace, is an offset of the i-th subspace, , , is determined by the historical experience of the technician, is a local feature segment split from the original data, for example, the fuel cell time series data is split into multiple subspaces according to the voltage trend, temperature trend, and voltage-temperature correlation.

[0042] In an implementation manner, one-dimensional time series data is converted into high-dimensional features containing causal relationships, which not only retains the advantage of processing long sequences of LSTM, but also breaks through the bottleneck of one-way information flow through a bidirectional mechanism, realizes high-precision identification of complex operation state monitoring modes with lower calculation cost, and provides interpretable operation state analysis basis, and is especially suitable for PEM fuel cell systems with extremely high safety and reliability requirements.

[0043] Based on the same inventive concept, the embodiments of the present application also provide an intelligent operation monitoring device of a PEM fuel cell. Referring to Figure 3 , Figure 3 is a framework diagram of an intelligent operation monitoring device of a PEM fuel cell provided by the embodiments of the present application, comprising: The data acquisition module is configured to acquire voltage data and temperature data of the PEM fuel cell. The preprocessing module is configured to preprocess the voltage data and the temperature data to obtain effective voltage data and effective temperature data. The first failure rate generation module is configured to perform data failure analysis on the effective voltage data and the effective temperature data to obtain a first failure rate. The two-dimensional image generation module is configured to perform two-dimensional data conversion on the effective voltage data and the effective temperature data by image encoding operation to obtain a two-dimensional image. The second failure rate generation module is configured to perform a flattening dimension reconstruction operation on the two-dimensional image to obtain a time series feature sequence, and substitutes the time series feature sequence into a lightweight long short-term network model to obtain a second failure rate. a detection result generation module configured to obtain a running state monitoring result according to the first failure rate and the second failure rate. The voltage data and the temperature data are data synchronously collected in a preset period before the current moment.

[0044] The intelligent running monitoring device of the PEM fuel cell provided by the embodiment of the present application eliminates the dimensional difference and the numerical range interference between the original data through preprocessing of the voltage data and the temperature data, and then adopts a double-path analysis strategy on the preprocessed data: first, a first failure rate is obtained through data failure analysis; second, a second failure rate is obtained by substituting a lightweight long short-term network model through image coding and de-flattening reconstruction operation, and finally a running state monitoring result is obtained according to the first failure rate and the second failure rate, so that the running state is evaluated from multiple dimensions, and the accuracy of running monitoring is improved.

[0045] In one embodiment, the preprocessing module comprises: a first normalization module configured to normalize the voltage data and the temperature data to obtain normalized voltage data and normalized temperature data; a first data segmentation module configured to divide the normalized voltage data and the normalized temperature data through a first preset sliding window to obtain a sub-voltage data set and a sub-temperature data set; a data replacement module configured to replace the average value of the sub-voltage data and the sub-temperature data in each window with the data points in the window to obtain effective voltage data and effective temperature data.

[0046] In one embodiment, the first failure rate generation module comprises: a second data segmentation module configured to divide the effective voltage data and the effective temperature data through a second preset sliding window to obtain a sub-effective voltage data set and a sub-effective temperature data set; a marking module configured to calculate the voltage slope and the temperature slope of the sub-effective voltage data and the sub-effective temperature data in each window, and if the absolute value of the voltage slope is greater than a preset voltage slope threshold or the absolute value of the temperature slope is greater than a preset temperature slope threshold, the data points in the window are marked to obtain a number of failure points; a first failure rate module configured to take the ratio of the number of failure points to the number of data points as the first failure rate.

[0047] In one embodiment, the two-dimensional image generation module comprises: a sorting module configured to arrange the effective voltage data and the effective temperature data in a time stamp order to obtain a one-dimensional sequence set; a second normalization module configured to normalize the one-dimensional data in the one-dimensional sequence set to obtain a normalized data set; The third normalization module is configured to perform a preset inverse cosine normalization operation on the normalized data in the normalized data set to obtain a two-dimensional feature matrix. The mapping module is configured to perform a preset pixel gray color value mapping on the numerical values in the two-dimensional feature matrix to obtain a two-dimensional image.

[0048] In one embodiment, the second failure rate generation module comprises: The compression module is configured to perform a compression matrix operation on the weight matrix of the long short-term memory network to obtain a low-rank approximation matrix. The replacement module is configured to replace the low-rank approximation matrix with the original weight matrix of the long short-term network to participate in calculation to obtain a lightweight long short-term network. The bidirectional processing module is configured to perform lightweight long short-term network processing on the time sequence feature sequence from the start end of the time sequence feature sequence in a forward direction and from the end of the time sequence feature sequence in a reverse direction step by step to obtain a forward state vector and a reverse state vector. The splicing module is configured to splice the forward state vector and the reverse state vector to obtain a fusion state vector. The second failure rate module is configured to perform a nonlinear transformation on the fusion feature vector by a preset monitoring formula to obtain a second failure rate.

[0049] The above describes one embodiment of the present application in detail, but the content is only the preferred embodiment of the present application, and cannot be considered as limiting the scope of the present application. Any equivalent changes and improvements made according to the scope of the present application should still belong to the patent coverage range of the present application.

Claims

1. A method for intelligent operation monitoring of a PEM fuel cell, characterized in that, The method includes: Simultaneously collect voltage and temperature data from the PEM fuel cell; The voltage data and temperature data are preprocessed respectively to obtain effective voltage data and effective temperature data; A first failure rate is obtained by performing data fault analysis on the effective voltage data and the effective temperature data. A two-dimensional image is obtained by performing two-dimensional data conversion on the effective voltage data and the effective temperature data through image encoding operations. The two-dimensional image is flattened and reconstructed to obtain a temporal feature sequence. The temporal feature sequence is then substituted into a lightweight long short-term network model to obtain a second failure rate. The operational status monitoring results are obtained based on the first failure rate and the second failure rate.

2. The intelligent operation monitoring method for a PEM fuel cell according to claim 1, characterized in that, Preprocessing the voltage data and temperature data to obtain effective voltage data and effective temperature data includes: The voltage data and the temperature data are normalized to obtain normalized voltage data and normalized temperature data; The normalized voltage data and the normalized temperature data are divided into sub-voltage datasets and sub-temperature datasets by using a first preset sliding window. The effective voltage data and effective temperature data are obtained by replacing the average value of the sub-voltage data and sub-temperature data within each window with the data points within the window.

3. The intelligent operation monitoring method for a PEM fuel cell according to claim 1, characterized in that, Data fault analysis of the effective voltage data and the effective temperature data yields a first fault rate, including: The effective voltage data and the effective temperature data are divided into sub-effective voltage datasets and sub-effective temperature datasets by using a second preset sliding window; Calculate the voltage slope and temperature slope of the sub-effective voltage data and sub-effective temperature data within each window. If the absolute value of the voltage slope is greater than the preset voltage slope threshold or the absolute value of the temperature slope is greater than the preset temperature slope threshold, then mark the data points within the window and obtain the number of fault points. The ratio of the total number of fault points to the total number of data points is taken as the first failure rate.

4. The intelligent operation monitoring method for a PEM fuel cell according to claim 1, characterized in that, The two-dimensional image is obtained by performing two-dimensional data conversion on the effective voltage data and the effective temperature data through image encoding operations, including: The effective voltage data and the effective temperature data are arranged in timestamp order to obtain a one-dimensional sequence set; Normalize the one-dimensional data in the one-dimensional sequence set to obtain a normalized dataset; A two-dimensional feature matrix is ​​obtained by performing a preset inverse cosine normalization operation on the normalized data in the normalized dataset. A two-dimensional image is obtained by mapping the values ​​in the two-dimensional feature matrix to preset pixel grayscale color values.

5. The intelligent operation monitoring method for a PEM fuel cell according to claim 1, characterized in that, The temporal feature sequence is obtained by flattening and reconstructing the dimensions of the two-dimensional image. Substituting the temporal feature sequence into a lightweight long short-term network model yields the second failure rate, which includes: A low-rank approximate matrix is ​​obtained by performing a compression matrix operation on the weight matrix of the Long Short-Term Memory network. The low-rank approximation matrix replaces the original weight matrix of the long short-term network in the calculation to obtain a lightweight long short-term network. The temporal feature sequence is processed step by step using a lightweight long short-term network to obtain a forward state vector and a reverse state vector, respectively, starting from the beginning of the temporal feature sequence in a forward direction and reversing from the end of the temporal feature sequence. The positive state vector and the negative state vector are concatenated to obtain the fused state vector; The second failure rate is obtained by performing a nonlinear transformation on the fused feature vector using a preset monitoring formula.

6. An intelligent operation monitoring device for a PEM fuel cell, characterized in that, The device includes: The data acquisition module is used to simultaneously acquire voltage and temperature data of the PEM fuel cell; The preprocessing module is used to preprocess the voltage data and the temperature data respectively to obtain effective voltage data and effective temperature data; The first failure rate generation module is used to perform data failure analysis on the effective voltage data and the effective temperature data to obtain a first failure rate. A two-dimensional image generation module is used to perform two-dimensional data conversion on the effective voltage data and the effective temperature data through image encoding operations to obtain a two-dimensional image. The second failure rate generation module is used to perform a flattening dimension reconstruction operation on the two-dimensional image to obtain a temporal feature sequence, and to substitute the temporal feature sequence into a lightweight long short-term network model to obtain a second failure rate. The detection result generation module is used to obtain the operation status monitoring results based on the first failure rate and the second failure rate.

7. The intelligent operation monitoring device for a PEM fuel cell according to claim 6, characterized in that, The preprocessing module includes: The first normalization module is used to normalize the voltage data and the temperature data to obtain normalized voltage data and normalized temperature data. The first data segmentation module is used to divide the normalized voltage data and the normalized temperature data into sub-voltage datasets and sub-temperature datasets through a first preset sliding window; The data replacement module is used to replace the data points in each window with the average value of the sub-voltage data and sub-temperature data to obtain the effective voltage data and effective temperature data.

8. The intelligent operation monitoring device for a PEM fuel cell according to claim 6, characterized in that, The first failure rate generation module includes: The second data segmentation module is used to divide the effective voltage data and the effective temperature data into sub-effective voltage datasets and sub-effective temperature datasets through a second preset sliding window; The marking module is used to calculate the voltage slope and temperature slope of sub-effective voltage data and sub-effective temperature data in each window. If the absolute value of the voltage slope is greater than the preset voltage slope threshold or the absolute value of the temperature slope is greater than the preset temperature slope threshold, the data points in the window are marked and the number of fault points is obtained. The first failure rate module is used to determine the ratio of the number of all failure points to the number of data points as the first failure rate.

9. The intelligent operation monitoring device for a PEM fuel cell according to claim 6, characterized in that, The 2D image generation module includes: The sorting module is used to arrange the effective voltage data and the effective temperature data in timestamp order to obtain a one-dimensional sequence set; The second normalization module is used to normalize the one-dimensional data in the one-dimensional sequence set to obtain a normalized dataset. The third normalization module is used to perform a preset inverse cosine normalization operation on the normalized data in the normalized dataset to obtain a two-dimensional feature matrix. The mapping module is used to map the values ​​in the two-dimensional feature matrix to preset pixel grayscale color values ​​to obtain a two-dimensional image.

10. The intelligent operation monitoring device for a PEM fuel cell according to claim 6, characterized in that, The second failure rate generation module includes: The compression module is used to perform matrix compression operations on the weight matrix of the Long Short-Term Memory network to obtain a low-rank approximate matrix. The replacement module is used to replace the original weight matrix of the long short-term network with the low-rank approximation matrix in the calculation to obtain a lightweight long short-term network. The bidirectional processing module is used to process the temporal feature sequence step by step using a lightweight long short-term network to obtain a forward state vector and a reverse state vector, respectively, from the beginning of the temporal feature sequence in a forward direction and from the end of the temporal feature sequence in a reverse direction. The splicing module is used to splice the positive state vector and the negative state vector to obtain a fused state vector; The second failure rate module is used to perform a nonlinear transformation on the fused feature vector using a preset monitoring formula to obtain a second failure rate.

Images