A Hydrogen Compressor Monitoring Method Based on Clustering Matrix and Dynamic Sliding Window

Abstract

The present invention belongs to the technical field of hydrogen compressor monitoring, and particularly relates to a hydrogen compressor monitoring method based on a clustering matrix and a dynamic sliding window. The monitoring method includes: for any thermodynamics parameter of the hydrogen compressor: obtaining the thermodynamics data of the current sliding window according to the window parameters of the current sliding window; calculating the statistical data of the current sliding window according to the thermodynamics data of the current sliding window; the window parameters, thermodynamics data and statistical data of the current sliding window jointly form the current sliding window information; according to the sliding window information of each thermodynamics parameter between the current monitoring moment and the previous monitoring moment, each clustering algorithm outputs the cluster result corresponding to the current monitoring moment; at the same time, calculating the weight coefficients of each clustering algorithm at the current monitoring moment; adding up the weight coefficients of the clustering algorithms corresponding to the same cluster result as the score of the corresponding cluster result, and taking the cluster result with the highest score as the operating state of the hydrogen compressor at the current monitoring moment. The present invention can improve the accuracy of monitoring the operating state of the hydrogen compressor.

Description

Technical Field

[0001] The present invention belongs to the technical field of hydrogen compressor monitoring, and particularly relates to a hydrogen compressor monitoring method based on a clustering matrix and a dynamic sliding window. Background Art

[0002] The hydrogen compressor (hereinafter referred to as "hydrogen compressor") is the core boosting equipment for hydrogen refueling stations and hydrogen filling, and is an important link in the wide application of hydrogen energy. The design pressure of the hydrogen compressor is high, up to more than 90 MPa at most, and the gas circuit, water circuit, and oil circuit inside the hydrogen compressor are complex and intertwined, and the number of various valves is numerous.

[0003] Due to the influence of assembly technology, transportation bumps, on-site operating environment, etc., even a hydrogen compressor that has passed the inspection at the time of leaving the factory may malfunction during actual operation; furthermore, as time goes by, the performance of the hydrogen compressor itself degrades, and internal parts age or are damaged, which will not only affect the operating efficiency of the hydrogen compressor, but also pose a threat to the personal safety of on-site technicians.

[0004] In addition to regularly overhauling and maintaining the key parts in the hydrogen compressor, technicians will also map the condition of the hydrogen compressor based on the exhaust pressure collected by the sensor during the operation of the hydrogen compressor through a certain clustering algorithm, that is, obtain the current working condition of the hydrogen compressor, so as to determine whether there is a problem with the current hydrogen compressor.

[0005] The existing technology is based on a single collected parameter, such as the exhaust pressure, and then uses a single clustering algorithm to determine which cluster the current condition of the hydrogen compressor will fall into, and the corresponding cluster result is output as the current working condition of the hydrogen compressor:

[0006] ① There are mutual influences among various parameters of the hydrogen compressor, which makes the existing technology based on this single thermodynamics parameter can only determine the working condition of the hydrogen compressor at the current moment, but cannot accurately predict the working condition of the hydrogen compressor in a future period of time.

[0007] ② When the clustering algorithm does not have the ability to generate cluster results autonomously (that is, the clustering algorithm is completely set by technicians for the cluster results, and the clustering algorithm will not generate new clustering clusters and corresponding cluster results): if a new working condition appears in the actual operation of the hydrogen compressor, and any set cluster result does not match the current working condition of the hydrogen compressor, then the current clustering algorithm will still output a cluster result set by technicians, that is, the flexibility of this clustering algorithm is poor and it cannot output the truly accurate working condition of the hydrogen compressor.

[0008] ③When the clustering algorithm has the ability to generate cluster results independently (that is, on the basis of the cluster results set by the technical personnel, the clustering algorithm will also independently generate new clustering clusters and corresponding cluster results): as the sample quantity of the acquisition parameters increases, the cluster results will also continuously increase, even showing an explosive growth. For the set cluster results, the technical personnel can directly take corresponding measures for the hydrogen compressor; however, for the new cluster results, the technical personnel need to verify and analyze them before determining whether to take measures for the hydrogen compressor and what measures to take; moreover, the verification and analysis by the technical personnel are lagging, and this greatly increases the workload of the technical personnel. In practical applications, after verification and analysis by the technical personnel, it is found that: some new cluster results are completely wrong; some new cluster results are correct, but they can completely belong to a set cluster result, and only a small number of new cluster results are accurate and different from the set cluster results.

[0009] Therefore, how to improve the accuracy of monitoring the operating state of the hydrogen compressor has become an urgent problem to be solved in the hydrogen compressor monitoring technology. Summary of the Invention

[0010] The object of the present invention is to overcome the deficiencies of the above-mentioned prior art and provide a hydrogen compressor monitoring method based on a clustering matrix and a dynamic sliding window, which can improve the accuracy of monitoring the operating state of the hydrogen compressor.

[0011] To achieve the above object, the present invention adopts the following technical solutions:

[0012] A hydrogen compressor monitoring method based on a clustering matrix and a dynamic sliding window includes the following steps:

[0013] S1, for any thermodynamics parameter of the hydrogen compressor: obtain the thermodynamics data of the current sliding window according to the window parameters of the current sliding window; the sliding windows are arranged in chronological order;

[0014] S2, calculate the statistical data of the current sliding window according to the thermodynamics data of the current sliding window; the window parameters, thermodynamics data, and statistical data of the current sliding window together form the current sliding window information;

[0015] S3, according to the sliding window information of each thermodynamics parameter between the current monitoring moment and the previous monitoring moment, each clustering algorithm outputs the cluster result corresponding to the current monitoring moment; at the same time, calculate the weight coefficient of each clustering algorithm at the current monitoring moment;

[0016] S4, accumulate the weight coefficients of the clustering algorithms corresponding to the same cluster result as the score of the corresponding cluster result, and then take the cluster result with the highest score as the operating state of the hydrogen compressor at the current monitoring moment.

[0017] Preferably, after S2, there is also S2': S2', calculating the window parameters of the next sliding window according to the statistical data of the current sliding window and the window parameters.

[0018] Preferably, when monitoring the current hydrogen compressor for the first time, or during the first operation after replacing more than 1 key component of the current hydrogen compressor, for any thermodynamics parameter of the current hydrogen compressor, a new monitoring cycle has started, and the window parameters of the first sliding window in the new monitoring cycle are initial values; the window parameters include the window length and the sliding step; there are several thermodynamics data arranged in time for only one thermodynamics parameter within each sliding window.

[0019] Preferably, the following content is also included in S2':

[0020] Denote the current sliding window with the window length of , where t is a positive integer, and the previous sliding window of the current sliding window is denoted as , then the window length of the next sliding window :

[0021] ,

[0022] ,

[0023] ,

[0024] ,

[0025] ,

[0026] where represents the window length adjustment amount of the current sliding window ; represents the window length adjustment amount of the next sliding window ; is the first adjustment coefficient; is the second adjustment coefficient; represents the weight coefficient of the i-th clustering algorithm, and a total of n clustering algorithms are used, that is, 1 ≤ i ≤ n; represents the intra-cluster entropy of the sliding window in the i-th clustering algorithm; represents the historical maximum entropy value of m sliding windows counted forward starting from the sliding window and including the sliding window , where m is a positive integer and m ≤ t; represents the mean vector of the sliding window , and the modulus of is the sliding window The difference from the average value of the thermodynamics data of the sliding window The direction of which points from the average value of the thermodynamics data of the sliding window to the average value of the thermodynamics data of the sliding window ; The average value of the thermodynamics data of the sliding window Denotes the 2-norm of ; Is the maximum value function; Denotes the entropy within the cluster in the i-th clustering algorithm of the sliding window ; Denotes the entropy within the cluster in the i-th clustering algorithm of the sliding window ; Denotes the entropy within the cluster in the corresponding clustering algorithm obtained by taking m sliding windows forward starting from the sliding window and including the sliding window during the process where i ranges from 1 to n; Denotes the covariance; Denotes the variance; Denotes the time constant of the hydrogen compressor system; Denotes the sampling period of the thermodynamics data in the current sliding window , which is determined by the acquisition frequency of the corresponding sensor; Denotes the maximum allowable drift rate of the thermodynamics parameters of the current sliding window ; Denotes the rated value of the thermodynamics parameters of the current sliding window .

[0027] Preferably, the sliding step of the next sliding window is: = where γ is the first constant parameter.

[0028] Preferably, the sliding step of the next sliding window is: =

[0029] ,

[0030] ,

[0031] ;

[0032] where Denotes the overlap ratio of the next sliding window ; Denotes the overlap ratio of the current sliding window ; Denotes the attenuation coefficient of the next sliding window ; represents the target overlap ratio; represents the total number of sliding windows that need to be transitioned currently; and are all known quantities determined by the operating state of the hydrogen compressor at the previous monitoring moment; represents including the next sliding window and within, the remaining number of sliding windows that need to be transitioned, , represents including the current sliding window and within, the remaining number of sliding windows that need to be transitioned.

[0033] Preferably, in S3, it further includes the following: n different clustering algorithms are adopted. Among these n clustering algorithms, some clustering algorithms do not have the ability to independently generate cluster results, and some other clustering algorithms have the ability to independently generate cluster results; the sliding window information of each thermodynamics parameter between the current monitoring moment and the previous monitoring moment is respectively sent into each clustering algorithm, and each clustering algorithm respectively outputs the cluster result corresponding to the hydrogen compressor at the current monitoring moment; meanwhile, the weight coefficients of each clustering algorithm at the current monitoring moment are calculated , 1 ≤ i ≤ n and i is a positive integer: If the current monitoring moment is the first monitoring moment within the current monitoring period, then the weight coefficients of each clustering algorithm at the current monitoring moment are obtained based on solving a multi-objective optimization function; if the current monitoring moment is not the first monitoring moment within the current monitoring period, then the weight coefficients of each clustering algorithm at the current monitoring moment are obtained based on the weight coefficients of each clustering algorithm at the previous monitoring moment.

[0034] Preferably, obtaining the weight coefficients of each clustering algorithm at the current monitoring moment based on solving a multi-objective optimization function further includes the following:

[0035] After constructing the multi-objective optimization function F and solving it, the values of the weight coefficients of each clustering algorithm at the current monitoring moment are obtained:

[0036] ,

[0037] ,

[0038] The constraint condition is: ,

[0039] wherein, is the second constant parameter; represents the weight coefficient corresponding to the i-th clustering algorithm; represents the sum of squared errors of the i-th clustering algorithm; represents the first weight matrix, which is a column vector composed of weight coefficients; represents the second weight matrix, which is the transpose of the first weight matrix and is a row vector; Represents the first weight matrix And the correlation matrix Of the quadratic form; Is the regularization parameter.

[0040] Preferably, based on the weight coefficients of each clustering algorithm at the previous monitoring moment, the weight coefficients of each clustering algorithm at the current monitoring moment are obtained, and the following also includes:

[0041] Denote the current monitoring moment as , the previous monitoring moment as , z≥2 and z is a positive integer, then the current monitoring moment The weight coefficient of the i-th clustering algorithm of Is:

[0042] ,

[0043] Among them, Represents the weight coefficient of the i-th clustering algorithm at the previous monitoring moment ; Represents the weight coefficient of the j-th clustering algorithm at the previous monitoring moment , 1≤j≤n and j is a positive integer; Represents the learning rate; Represents the global distribution at the previous monitoring moment; Represents the distribution of the i-th clustering algorithm at the previous monitoring moment; Represents the distribution of the j-th clustering algorithm at the previous monitoring moment; Represents the JS divergence; Represents the distribution And the global distribution Of the JS divergence; Represents the distribution And the global distribution Of the JS divergence.

[0044] Preferably, after S4, S5 is also included: S5, if the cluster result obtained by the clustering algorithm with the ability to autonomously generate cluster results does not become the cluster result with the highest score, then the cluster result obtained by the clustering algorithm with the ability to autonomously generate cluster results is recorded as a set to be verified. After technicians regularly verify and analyze the set to be verified, if the cluster result obtained by a certain clustering algorithm in the set to be verified is continuously considered correct by technicians for y times and is different from the cluster result set by technicians, then technicians increase the weight coefficient of the corresponding clustering algorithm.

[0045] The beneficial effects of the present invention are as follows:

[0046] (1) The hydrogen compressor monitoring method of the present invention can improve the accuracy of monitoring the operating state of the hydrogen compressor.

[0047] (2) A single parameter cannot effectively monitor and warn the device. The hydrogen compressor monitoring method of the present invention does not directly process the single type of data collected by the sensor, but takes the sliding window as the unit of collecting data, and clusters the sliding window information between two adjacent monitoring times. For any thermodynamics parameter, the thermodynamics data in a sliding window is time-sequential, and there is also time-sequentiality between adjacent sliding windows (new sliding windows can only be born as time goes by); there is also partial overlap of thermodynamics data between adjacent sliding windows, which enables the clustering algorithm to well mine and learn the potential relationships and correlations between adjacent sliding windows when processing the time-sequential sliding window information, and improve the accuracy of the cluster results obtained by each clustering algorithm in the present invention.

[0048] (3) In the hydrogen compressor monitoring method of the present invention, a sliding window only contains one type of thermodynamics parameter. The sliding windows of different thermodynamics parameters within the same time period overlap in time because the time of collecting data by the sensor overlaps, so that the sliding windows also overlap in time. The clustering algorithm in the present invention is not based on the isolated thermodynamics data at a certain moment; but based on this time-sequential, time-related and overlapping thermodynamics data sliding window information, and there is also time overlap between the sliding windows of different thermodynamics parameters. The clustering algorithm will also mine and learn the potential relationships and correlations of the sliding window information of these different thermodynamics parameters with time overlap, which enables each clustering algorithm in the present invention to more accurately predict the operating conditions of the current hydrogen compressor in the next period of time, thereby improving the accuracy of the cluster results output by each clustering algorithm; this also greatly improves the accuracy of the new clustering clusters and the corresponding cluster results obtained by the clustering algorithm with the ability to autonomously generate cluster results.

[0049] (4) The window length and sliding step of the sliding window are directly related to the sample data input into each clustering algorithm, and indirectly related to the accuracy of the cluster results output by each clustering algorithm. In each monitoring cycle of the present invention, the window length and sliding step are automatically adjusted on the basis of the previous sliding window according to the previous several sliding windows and the intra-cluster entropy obtained by each clustering algorithm based on the previous several sliding windows, without manual intervention. Technicians only need to set the window parameters (i.e., initial values) of the first sliding window in each monitoring cycle, which greatly reduces the labor cost.

[0050] (5) During the entire monitoring process of the present invention, the adjustment of the window length and sliding step of the sliding window not only has a high degree of automation, but also balances the computational overhead and storage resource occupancy during the process of processing the sliding window information. At the same time, it enables the clustering algorithm to better mine and learn the potential relationships and associations between sliding windows, allowing each clustering algorithm to better process the sliding window information and effectively perceive the long-term slowly deteriorating monitoring object.

[0051] (6) The monitoring method of the present invention will automatically update the monitoring cycle, which is reflected in that when initially monitoring the current hydrogen compressor or during the first operation after replacing one or more key parts of the current hydrogen compressor, for any thermodynamics parameter of the current hydrogen compressor, a new monitoring cycle has started. Even for hydrogen compressors of the same model that have passed the pre-factory test, there are certain pre-factory test data. However, the period during which a hydrogen compressor is officially put into operation after being transported to the site and installed is the time period monitored by the monitoring method of the present invention, that is, when initially monitoring a certain hydrogen compressor. Therefore, before initially monitoring a certain hydrogen compressor, the internal parts of the hydrogen compressor may still be damaged or there may be installation mismatches due to transportation and on-site installation. The same applies to the first operation after replacing one or more key parts of the current hydrogen compressor. Both of these situations mean that the hydrogen compressor as a whole needs to go through a new running-in period, stable period, and deterioration period again. And within each monitoring cycle, the window length and sliding step are based on the previous sliding window. Therefore, it is inevitable that the window length and sliding step of the new sliding window in these two situations cannot be obtained based on the previous sliding window. So we need to let the hydrogen compressor in these two situations enter a new monitoring cycle.

[0052] (7) The update of the window length and sliding step of the sliding window is related to the frequency of sample input and the change in the amount of sample data for the subsequent clustering algorithm. The monitoring method of the present invention automatically updates the monitoring cycle and automatically adjusts the window length and sliding step of the sliding window within each monitoring cycle to make the frequency of sample input and the amount of sample data adapt to the subsequent processing process of the clustering algorithm, further improving the efficiency of the entire monitoring process and the accuracy of the cluster results output by the clustering algorithm. Therefore, the monitoring method of the present invention can better adapt to each hydrogen compressor it monitors.

[0053] (8) The hydrogen compressor monitoring method of the present invention does not use a single clustering algorithm to monitor the operating status of the hydrogen compressor, but uses multiple clustering algorithms for joint monitoring. These clustering algorithms include not only clustering algorithms that have the ability to autonomously generate cluster results, but also clustering algorithms that do not have the ability to autonomously generate cluster results. The cluster results obtained by each of these clustering algorithms may be partially the same or completely different. However, the present invention will ultimately use the cluster result with the highest score as the operating status of the hydrogen compressor at the current monitoring time. The cluster result with the highest score is also the cluster result of most clustering algorithms. After random inspection by technical personnel, they are basically the operating status of the hydrogen compressor at the current monitoring time.

[0054] (9) The score of the cluster result is obtained by accumulating the weight coefficients of the clustering algorithms corresponding to the same cluster result, and the weight coefficients in the monitoring method of the present invention also change with the monitoring cycle and the monitoring time within each monitoring cycle: the weight coefficients of each clustering algorithm at the first monitoring time in each monitoring cycle are determined by solving the multi-objective optimization function; and the weight coefficients of each clustering algorithm at the remaining monitoring times in each monitoring cycle need to be obtained based on the weight coefficients of each clustering algorithm at the previous monitoring time. This ensures that the weight coefficients of each clustering algorithm will not change suddenly while changing with the monitoring time; and it has been verified by technical personnel that only when a clustering algorithm obtains several cluster results in succession that are all the operating status of the hydrogen compressor at the corresponding monitoring time, the weight coefficient of the clustering algorithm will gradually increase.

[0055] (10) The hydrogen compressor monitoring method of the present invention also takes into account a special case with an extremely small probability: if the clustering algorithm with the ability to autonomously generate cluster results obtains a new cluster result corresponding to the current monitoring time, and the weight coefficient of the clustering algorithm that generates this new cluster result is small, then this new cluster result cannot be used as the operating status of the hydrogen compressor at the current monitoring time; and the content of S5 is adopted accordingly. In this way, on the one hand, it can be avoided that the clustering algorithm with the ability to autonomously generate cluster results will generate multiple new cluster results without restraint to interfere with the technicians' evaluation of the operating status of the hydrogen compressor, greatly reducing the workload of the technicians. On the other hand, it can also ensure through a small amount of manual intervention that when the hydrogen compressor really has a situation other than the cluster result preset by the technicians, the clustering algorithm with a high weight coefficient and the ability to autonomously generate cluster results can accurately determine the operating status of the hydrogen compressor at the current monitoring time.

[0056] (11) In the hydrogen compressor monitoring method of the present invention, the operating state of the hydrogen compressor at the current monitoring moment can finally be accurately determined. The operating state of the hydrogen compressor not only includes the operating conditions of the hydrogen compressor at the current monitoring moment, but also includes the prediction of the operating conditions of the current hydrogen compressor within a certain period in the future. Therefore, the number of cluster results preset by technicians can be more. For example, the determination of the operating conditions of the hydrogen compressor at the current monitoring moment in multiple cluster results is the same, but the prediction of the operating conditions of the current hydrogen compressor within a certain period in the future is different. This is directly different from the prior art in that only the operating conditions of the hydrogen compressor at the current monitoring moment can be determined by using a single thermodynamic parameter, reducing the risk of misjudgment in the prior art that may be caused by technicians presetting too many cluster results. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a flowchart of a hydrogen compressor monitoring method based on a clustering matrix and a dynamic sliding window according to the present invention;

[0058] Figures 2a to 2c are the corresponding clustering results obtained by using the DBSCAN clustering algorithm according to different thermodynamic parameters;

[0059] Figures 3a to 3c are the corresponding clustering results obtained by using the OPTICS clustering algorithm according to different thermodynamic parameters;

[0060] Figures 4a to 4c are the corresponding clustering results obtained by using the GMM clustering algorithm according to different thermodynamic parameters;

[0061] Figures 5a to 5c are the corresponding clustering results obtained by using the hydrogen compressor monitoring method of the present invention according to different thermodynamic parameters. DETAILED DESCRIPTION OF THE INVENTION

[0062] To make the technical solutions of the present invention clearer and more definite, the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All solutions obtained by equivalent replacement and conventional reasoning of the technical features of the technical solutions of the present invention by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0063] As Figure 1 shown, it is a flowchart of a hydrogen compressor monitoring method based on a clustering matrix and a dynamic sliding window according to the present invention, including the following steps:

[0064] S1, for any thermodynamic parameter of the hydrogen compressor: obtain the thermodynamic data of the current sliding window according to the window parameters of the current sliding window; the sliding windows are arranged in chronological order;

[0065] S2. Calculate the statistical data of the current sliding window based on the thermodynamics data of the current sliding window; the window parameters, thermodynamics data, and statistical data of the current sliding window together constitute the current sliding window information.

[0066] S3. Based on the sliding window information of each thermodynamics parameter between the current monitoring moment and the previous monitoring moment, each clustering algorithm outputs the cluster result corresponding to the current monitoring moment; meanwhile, calculate the weight coefficients of each clustering algorithm at the current monitoring moment.

[0067] S4. Accumulate the weight coefficients of the clustering algorithms corresponding to the same cluster result as the score of the corresponding cluster result, and then take the cluster result with the highest score as the operating state of the hydrogen compressor at the current monitoring moment.

[0068] After S2, it also includes S2':

[0069] S2'. Calculate the window parameters of the next sliding window based on the statistical data and window parameters of the current sliding window.

[0070] In S1, it also includes the following content:

[0071] The thermodynamics parameters of the hydrogen compressor include: inlet pressure, exhaust pressure, inter-stage pressure, compressor flow rate, inlet temperature, temperature before the first-stage exhaust cooling, temperature after the first-stage exhaust cooling, temperature before the second-stage exhaust cooling, temperature after the second-stage exhaust cooling, inlet water temperature, return water temperature, lubricating oil pressure, lubricating oil temperature, first-stage cylinder pressure, second-stage cylinder pressure, first-stage bearing temperature, second-stage bearing temperature, cooler temperature.

[0072] The key parts of the hydrogen compressor include diaphragm assembly, valve assembly, crankshaft connecting rod mechanism, sealing system, cooling system, drive motor and control system, etc. Among them, the diaphragm assembly is the core power transmission component, and any changes in its diaphragm material, thickness, number of layers, etc. are regarded as replacing the diaphragm assembly; the valve assembly includes suction valve / exhaust valve, and any changes in valve disc material, spring stiffness, valve clearance, etc. are regarded as replacing the valve assembly; the crankshaft connecting rod mechanism is the core of power conversion, and any changes in its clearance, counterweight, etc. are regarded as replacing the crankshaft connecting rod mechanism; the sealing system includes static seal and dynamic seal, and any changes in its packing material and size, etc. are regarded as replacing the sealing system; any changes in the properties of the cooling medium in the cooling system are regarded as replacing the cooling system; any changes in motor efficiency, power, PID control parameters, etc. in the drive motor and control system are regarded as replacing the drive motor and control system.

[0073] When the current hydrogen compressor is monitored for the first time, or during the first run after replacing one or more key components of the current hydrogen compressor, for any thermodynamic parameter of the current hydrogen compressor, a new monitoring cycle is entered, and the window parameters of the first sliding window in the new monitoring cycle are all initial values. The window parameters include the window length and the sliding step. The initial values of the window parameters are set by technicians, and the initial window values for different thermodynamic parameters can be different.

[0074] Except for the above situations, no matter how many times the current hydrogen compressor is restarted after shutdown, for any thermodynamic parameter of the hydrogen compressor, it is in the same monitoring cycle. In the same monitoring cycle, for a certain thermodynamic parameter, assuming that the last sliding window during the previous run of the current hydrogen compressor is the 51st sliding window in the current cycle, then the first sliding window after the current hydrogen compressor is restarted is the 52nd sliding window in the current cycle.

[0075] Each sliding window contains several thermodynamic data arranged in time for only one thermodynamic parameter. There is an overlap of several thermodynamic data between adjacent sliding windows of the same thermodynamic parameter.

[0076] The sliding direction of the sliding window is unique and synchronized with the time sequence; the sliding window slides one sliding step each time, and the sliding step is a time quantity. After the sliding window completes one slide, it becomes a new sliding window. Each time the sliding window slides, its sliding step is the sliding step in the window parameters of the current sliding window.

[0077] For the same thermodynamic parameter: the window length and the window sliding step directly affect the number of overlapping thermodynamic data between adjacent sliding windows. The window length directly determines the number of thermodynamic data in the sliding window.

[0078] Thermodynamic data is collected regularly by the corresponding sensors, and the acquisition frequency of the sensors is generally fixed.

[0079] The following is also included in S2:

[0080] The statistical data of the current sliding window includes: the average value of the thermodynamic data, the variance of the thermodynamic data, the standard deviation of the thermodynamic data, the median of the thermodynamic data, the first quartile of the thermodynamic data, the third quartile of the thermodynamic data, the interquartile range of the thermodynamic data, the skewness of the thermodynamic data, the kurtosis of the thermodynamic data, the maximum value of the thermodynamic data, the minimum value of the thermodynamic data, and the coefficient of variation of the thermodynamic data.

[0081] The following is also included in S2´:

[0082] Denote the current sliding window with a window length of , where \(t\) is a positive integer, and the current sliding window The previous sliding window is denoted as , then the window length of the next sliding window is: :

[0083] ,

[0084] ,

[0085] ,

[0086] ,

[0087] ;

[0088] Among them, represents the window length adjustment amount of the current sliding window ; represents the window length adjustment amount of the next sliding window ; is the first adjustment coefficient; is the second adjustment coefficient; represents the weight coefficient of the \(i\)-th clustering algorithm. In the present invention, a total of \(n\) clustering algorithms are adopted, and \(1\leq i\leq n\); represents the within-cluster entropy of the sliding window in the \(i\)-th clustering algorithm; represents the historical maximum entropy value of \(m\) sliding windows counted forward starting from the sliding window (including the sliding window ), where \(m\) is a positive integer and \(m\leq t\); represents the mean vector of the sliding window , The modulus of is the difference between the sliding window and the average value of the thermodynamics data of the sliding window The direction of points from the average value of the thermodynamics data of the sliding window to the average value of the thermodynamics data of the sliding window represents the 2-norm of is the maximum value function; represents the within-cluster entropy of the sliding window in the \(i\)-th clustering algorithm, represents the within-cluster entropy of the sliding window in the \(i\)-th clustering algorithm; represents starting from the sliding window (including the sliding window m sliding windows are counted forward, and the intra-cluster entropy in the corresponding clustering algorithm obtained during the process of taking i from 1 to n; represents the covariance; represents the variance; represents the time constant of the hydrogen compressor system; represents the current sliding window The sampling period of the thermodynamics data in it is determined by the acquisition frequency of the corresponding sensor; represents the current sliding window The maximum allowable drift rate of the thermodynamics parameters of is set by technicians according to experience; represents the current sliding window The rated value of the thermodynamics parameters of.

[0089] If 1 = t < m, that is, starting from the sliding window (including the sliding window ), there are only t sliding windows counted forward, then = ; if 2 = t < m, then = ; if 3 = t < m, then = .

[0090] Optionally, calculate the sliding step size of the next sliding window as: is: = , where γ is the first constant parameter, defined by the technician himself. In this embodiment, γ = 0.2.

[0091] Optionally, calculate the sliding step size of the next sliding window as: :

[0092] ,

[0093] ,

[0094] ;

[0095] Among them, represents the overlap ratio of the next sliding window ; represents the overlap ratio of the current sliding window ; represents the attenuation coefficient of the next sliding window ; represents the target overlap ratio; represents the total number of sliding windows that need to be transitioned currently; and All are determined by the operating state of the hydrogen compressor at the previous monitoring moment and are all known quantities. In the present invention, those skilled in the art will pre-divide several levels of target overlap ratios and the total number of sliding windows to be transitioned corresponding thereto according to the operating state of the hydrogen compressor; Indicates including the next sliding window and the remaining number of sliding windows to be transitioned; , Indicates including the current sliding window and the remaining number of sliding windows to be transitioned.

[0096] In the present invention, in the process of determining "the sliding step length of the next sliding window ": Based on the smooth transition algorithm of state transition, a state transition equation is constructed, that is and ; Timely adjustment of the sliding step length of the next sliding window is to timely adjust the number of thermodynamics data overlapping between adjacent sliding windows; and through the transition of continuously adjusting several sliding step lengths, that is, several sliding windows gradually change the sliding step length, a huge adjustment of the sliding step length in stages can be achieved, avoiding affecting the mining and learning of the potential relationship and correlation between adjacent sliding windows by the subsequent clustering algorithm due to too large a change in the sliding step length between adjacent sliding windows. ;

[0097] The sliding window is the first sliding window.

[0098] In this example, if the average value of the thermodynamics data of the sliding window is greater than the average value of the thermodynamics data of the sliding window , then the direction of

[0099] is positive, otherwise it is negative. During each sliding of the sliding window, the window length is adjusted, that is, when the sliding window completes one sliding, the window length is also adjusted, and the window length of the next sliding window formed is

[0100] . Only by adjusting the right boundary of the sliding window can the window length of the sliding window be adjusted; the left boundary of the sliding window does not need to be adjusted, and the distance between the left boundaries of adjacent sliding windows is a sliding step length.

[0101] In S3, the following content is also included:

[0102] Adopt n different clustering algorithms. Among these n clustering algorithms, some do not have the ability to autonomously generate cluster results, while others do. Send the sliding window information of each thermodynamic parameter between the current monitoring moment and the previous monitoring moment into each clustering algorithm respectively, and each clustering algorithm outputs the cluster results corresponding to the hydrogen compressor at the current monitoring moment.

[0103] At the same time, calculate the weight coefficients of each clustering algorithm at the current monitoring moment , 1 ≤ i ≤ n and i is a positive integer: If the current monitoring moment is the first monitoring moment within the current monitoring period, then obtain the weight coefficients of each clustering algorithm at the current monitoring moment by solving the multi-objective optimization function; If the current monitoring moment is not the first monitoring moment within the current monitoring period, then obtain the weight coefficients of each clustering algorithm at the current monitoring moment based on the weight coefficients of each clustering algorithm at the previous monitoring moment.

[0104] In this embodiment, "send the sliding window information of each thermodynamic parameter between the current monitoring moment and the previous monitoring moment into each clustering algorithm respectively", and this sliding window refers to the sliding window whose right boundary is between the current monitoring moment and the previous monitoring moment.

[0105] In this embodiment, 10 different clustering algorithms are adopted, namely AP clustering algorithm (Affinity Propagation clustering algorithm), Agglomerative clustering algorithm (Agglomerative hierarchical clustering algorithm), BIRCH clustering algorithm (Balanced Iterative Reducing and Clustering using Hierarchies algorithm), DBSCAN clustering algorithm (Density-Based Spatial Clustering of Applications with Noise algorithm), GMM clustering algorithm (Gaussian Mixture Model clustering algorithm), Divisive clustering algorithm (Divisive hierarchical clustering algorithm), Kmeans++ clustering algorithm (K-Means++ algorithm), MeanShift clustering algorithm (Mean Shift algorithm), OPTICS clustering algorithm (Ordering Points To Identify the Clustering Structure algorithm), Spectral clustering algorithm (Spectral clustering algorithm); corresponding to the weight coefficients in turn ~ 。

[0106] Among them, the Agglomerative clustering algorithm, GMM clustering algorithm, Divisive clustering algorithm, Kmeans++ clustering algorithm and Spectral clustering algorithm do not have the ability to autonomously generate cluster results; the AP clustering algorithm, BIRCH clustering algorithm, DBSCAN clustering algorithm, MeanShift clustering algorithm and OPTICS clustering algorithm have the ability to autonomously generate cluster results.

[0107] "Obtain the weight coefficients of each clustering algorithm at the current monitoring moment by solving the multi-objective optimization function" also includes the following content:

[0108] Based on the Lyapunov stability theory, after constructing the multi-objective optimization function F and solving it, the values of the weight coefficients of each clustering algorithm at the current monitoring moment are obtained:

[0109] ,

[0110] ,

[0111] The constraint conditions are: ,

[0112] Among them, is the second constant parameter. In this embodiment, = 0.001; represents the weight coefficient corresponding to the i-th clustering algorithm; represents the sum of squared errors of the i-th clustering algorithm; represents the first weight matrix, which is also a column vector composed of weight coefficients; represents the second weight matrix, which is the transpose of the first weight matrix and is also a row vector; represents the first weight matrix and the quadratic form of the correlation matrix ; is the regularization parameter.

[0113] The correlation matrix is used to describe the relationship between different clustering algorithms; the regularization parameter is set by the technical personnel and is used to control the influence of the correlation between clustering algorithms.

[0114] In this embodiment, for the multi-objective optimization function F, the CVXPY tool is used to convert it into a convex optimization, and then the built-in solvers ECOS and SCS are called for solution.

[0115] "Obtaining the weight coefficients of each clustering algorithm at the current monitoring moment based on the weight coefficients of each clustering algorithm at the previous monitoring moment" also includes the following content:

[0116] Denote the current monitoring moment as , the previous monitoring moment as , z ≥ 2 and z is a positive integer. Then the weight coefficient of the i-th clustering algorithm at the current monitoring moment is:

[0117] ,

[0118] Among them, represents the weight coefficient of the i-th clustering algorithm at the previous monitoring moment ; Indicates the previous monitoring moment which is the weight coefficient of the j-th clustering algorithm, where 1 ≤ j ≤ n and j is a positive integer; Indicates the learning rate, which is set by the technician and used to control the update speed of the algorithm weight coefficient. When it increases, it means that the algorithm adjusts the weight coefficient more actively to adapt to the distribution faster; Indicates the global distribution at the previous monitoring moment. The global distribution at the previous monitoring moment can be obtained comprehensively based on the thermodynamic data recorded during the current monitoring period and the weight coefficients of each algorithm, and it is a known quantity; Indicates the distribution of the i-th clustering algorithm at the previous monitoring moment; Indicates the distribution of the j-th clustering algorithm at the previous monitoring moment; Indicates the distribution of the j-th clustering algorithm at the previous monitoring moment; Indicates the JS divergence; Indicates the distribution and the global distribution of the JS divergence, which is used to measure the difference between and . The more similar the distribution is to the global distribution , the larger the value is; Indicates the distribution and the global distribution of the JS divergence, which is used to measure the difference between and . The more similar the distribution is to the global distribution , the larger the value is.

[0119] After S4, it also includes S5:

[0120] S5. If the cluster result obtained by the clustering algorithm with the ability to autonomously generate cluster results does not become the cluster result with the highest score, then the cluster result obtained by the clustering algorithm with the ability to autonomously generate cluster results is recorded as a set to be verified. After the technician regularly checks and analyzes the set to be verified, if the cluster result obtained by a certain clustering algorithm in the set to be verified is considered correct by the technician for y consecutive times and is different from the cluster result set by the technician, then the technician increases the weight coefficient of the corresponding clustering algorithm.

[0121] In this embodiment, y = 5.

[0122] In the prior art, a single type of data collected by a sensor is directly processed. Since the data at the current time point is processed through clustering to obtain a cluster result, and the corresponding cluster result is output as the current working condition of the hydrogen compressor. Therefore, the prior art can generally only be used to determine the current working condition of a certain hydrogen compressor; however, the determination of the current working condition of a certain hydrogen compressor is also affected by the clustering algorithm selected. This impact will directly manifest in reducing the accuracy of judging the current working condition of the hydrogen compressor, which we have already described in the background art and will not be elaborated here.

[0123] Suppose the set cluster result 1 by the technician is "the hydrogen compressor is running normally continuously", and the cluster result 2 is "the hydrogen compressor is currently running normally, but in a low-pressure operating state". It can be seen that cluster result 1 not only contains the determination of the current working condition "the hydrogen compressor is running normally", but also contains the prediction of the working condition of the current hydrogen compressor in a period of time in the future "running normally continuously"; in cluster result 2, it not only contains the determination of the current working condition of the hydrogen compressor "the hydrogen compressor is running normally continuously", but also contains the prediction of the working condition of the current hydrogen compressor in a period of time in the future "in a low-pressure operating state". However, in the clustering algorithms of the prior art (regardless of whether they have the ability to generate cluster results independently), since the data at the current time point is processed through clustering, it cannot well predict the working condition of the current hydrogen compressor in a period of time in the future (for example, the current hydrogen compressor is indeed "in a low-pressure operating state" in a period of time in the future, but the prior art determines that it will "run normally continuously" in a period of time in the future). Then, even if the operating state of the current hydrogen compressor should more appropriately fall into the cluster where cluster result 2 is located, the clustering algorithm of the prior art will make it fall into cluster result 1.

[0124] Single parameters cannot effectively realize the monitoring and early warning of equipment. The hydrogen compressor monitoring method of the present invention does not directly process a single type of data collected by the sensor, but takes the sliding window as the unit for collecting data, and performs clustering processing on the sliding window information between two adjacent monitoring moments. For any thermodynamics parameter, the thermodynamics data in a sliding window is sequential, and there is also sequentiality between adjacent sliding windows (a new sliding window can only be born as time goes by); there is also partial overlap of thermodynamics data between adjacent sliding windows, which enables the clustering algorithm to well mine and learn the potential relationships and correlations between adjacent sliding windows when processing the sequential sliding window information, and improve the accuracy of the cluster results obtained by each clustering algorithm in the present invention.

[0125] In the hydrogen compressor monitoring method of the present invention, only one type of thermodynamics parameter is included in a sliding window. Since the time for the sensor to collect data for different thermodynamics parameters overlaps within the same time period, the sliding windows also overlap in time. The clustering algorithm in the present invention is not based on the isothermal thermodynamic data at a certain moment; instead, it is based on the sliding window information of such time-series, temporally related, and overlapping thermodynamic data. Moreover, there is also a time overlap between the sliding windows of different thermodynamics parameters. The clustering algorithm will also explore and learn the potential relationships and associations between the sliding window information of these different thermodynamics parameters with time overlap. This enables each clustering algorithm in the present invention to accurately predict the operating conditions of the current hydrogen compressor in the next period of time, thereby improving the accuracy of the cluster results output by each clustering algorithm; this also greatly improves the accuracy of the new clustering clusters and the corresponding cluster results obtained by the clustering algorithm with the ability to autonomously generate cluster results.

[0126] As can be seen from the above analysis, the window length and sliding step of the sliding window are directly related to the sample data input into each clustering algorithm and indirectly related to the accuracy of the cluster results output by each clustering algorithm. In each monitoring cycle of the present invention, the window length and sliding step are automatically adjusted based on the previous sliding window, the previous several sliding windows, and the intra-cluster entropy obtained by each clustering algorithm based on the previous several sliding windows, without manual intervention. Technicians only need to set the window parameters (i.e., the initial values) of the first sliding window in each monitoring cycle, which greatly reduces the labor cost. If the window length is too long or the sliding step is too short, it will lead to a significant increase in the computational cost during the process of the clustering algorithm processing the sliding window information and occupy a large amount of storage resources; while if the window length is too short or the sliding step is too long, although the computational cost and the occupation of storage resources are reduced, the overlapping thermodynamic data between adjacent sliding windows is also greatly reduced or even zero, and the distribution of the sliding windows with time overlap between the sliding windows of different thermodynamics parameters will also change. This is very unfavorable for the clustering algorithm to explore and learn the potential relationships and associations between the sliding windows, and even affects the accuracy of the cluster results output by each clustering algorithm. Therefore, the adjustment of the window length and sliding step of the sliding window in the entire monitoring process of the present invention not only has a high degree of automation, but also balances the computational cost and the occupation of storage resources during the process of processing the sliding window information, and enables the clustering algorithm to better explore and learn the potential relationships and associations between the sliding windows, allowing each clustering algorithm to better process the sliding window information and effectively perceive the long-term slowly deteriorating monitoring object.

[0127] The monitoring method of the present invention automatically updates the monitoring period, which is reflected in that when monitoring the current hydrogen compressor for the first time or during the first operation after replacing one or more key parts of the current hydrogen compressor, for any thermodynamics parameter of the current hydrogen compressor, a new monitoring period has started. Even for hydrogen compressors of the same model that have passed the pre-factory test and have certain pre-factory test data, after a hydrogen compressor is transported and installed at the site and starts formal operation, this is the period monitored by the monitoring method of the present invention, that is, when monitoring a certain hydrogen compressor for the first time. Therefore, before monitoring a certain hydrogen compressor for the first time, some parts inside the hydrogen compressor may still be damaged or there may be installation mismatches due to transportation and on-site installation; the same is true for the first operation after replacing one or more key parts of the current hydrogen compressor; both of these situations mean that the entire hydrogen compressor needs to go through the running-in period, the stable period, and the deterioration period again. And within each monitoring period, the window length and the sliding step length are based on the previous sliding window. Therefore, the window length and the sliding step length of the new sliding window in these two situations cannot be obtained based on the previous sliding window. So, we need to let the hydrogen compressors in these two situations enter a new monitoring period.

[0128] Based on the above analysis, it can also be seen that the update of the window length and the sliding step length of the sliding window is related to the frequency of sample input and the change in the amount of sample data for the subsequent clustering algorithm. The monitoring method of the present invention makes the frequency of sample input and the amount of sample data adapt to the subsequent clustering algorithm processing process by automatically updating the monitoring period and automatically adjusting the window length and the sliding step length within each monitoring period, further improving the efficiency of the entire monitoring process and the accuracy of the cluster results output by the clustering algorithm; therefore, the monitoring method of the present invention can better adapt to each hydrogen compressor it monitors. For easy understanding, for example: after entering a new monitoring period, in order to better understand the situation of the current hydrogen compressor within the current monitoring period, the monitoring method of the present invention needs to frequently perform sample input and increase the amount of sample data input; after verification by technical personnel, within a certain period of time after updating the monitoring period (i.e., the running-in period), the window length of the sliding window of the present invention indeed increases slowly, and the sliding step length of the sliding window is short, to meet the need to frequently perform sample input and increase the amount of sample data input within the running-in period of the current monitoring period; during the middle period of the new monitoring period (i.e., the stable period), the window length and the sliding step length of the sliding window both increase significantly, and during the stable period of the current monitoring period, there is no need to frequently perform sample input or input a large amount of sample data (reflected in the reduction of overlapping thermodynamic data between adjacent sliding windows and the increase of the sliding step length of the sliding window); subsequently, as the clustering algorithm determines the operating condition of the hydrogen compressor, the frequency of sample input and the amount of sample data input will increase again because the hydrogen compressor enters the deterioration period of the current monitoring period, which will not be elaborated here.

[0129] The hydrogen compressor monitoring method of the present invention does not use a single clustering algorithm to monitor the operating state of the hydrogen compressor, but uses multiple clustering algorithms for joint monitoring. Moreover, these clustering algorithms not only include clustering algorithms with the ability to autonomously generate cluster results, but also include clustering algorithms without the ability to autonomously generate cluster results; the cluster results obtained by these clustering algorithms may partially be the same or may all be different. However, the present invention will ultimately use the cluster result with the highest score as the operating state of the hydrogen compressor at the current monitoring moment. The cluster result with the highest score is also the cluster result of most clustering algorithms. After spot checks by technicians, it is basically the operating state of the hydrogen compressor at the current monitoring moment.

[0130] The score of the cluster result is obtained by accumulating the weight coefficients of the clustering algorithms corresponding to the same cluster result. Moreover, the weight coefficients in the monitoring method of the present invention also change with the monitoring period and the monitoring moments within each monitoring period: the weight coefficients of each clustering algorithm at the first monitoring moment within each monitoring period are determined by solving a multi-objective optimization function; while the weight coefficients of each clustering algorithm at the remaining monitoring moments within each monitoring period need to be obtained based on the weight coefficients of each clustering algorithm at the previous monitoring moment. This makes the weight coefficients of each clustering algorithm not mutate while changing with the monitoring moment; and after verification by technicians, only when a certain clustering algorithm continuously obtains several cluster results that are all the operating state of the hydrogen compressor at the corresponding monitoring moment, will the weight coefficient of this clustering algorithm gradually increase.

[0131] Considering the special case with extremely low probability: for a clustering algorithm capable of autonomously generating cluster results, if the cluster result corresponding to the current monitoring moment is a new cluster result and the weight coefficient of the clustering algorithm that generates this new cluster result is small, then this new cluster result cannot be used as the operating state of the hydrogen compressor at the current monitoring moment. Therefore, in the present invention, the cluster results obtained by all clustering algorithms capable of autonomously generating cluster results are recorded as a set to be verified. That is, in S5, after technicians regularly verify and analyze the set to be verified, if the cluster results obtained by a certain clustering algorithm in the set to be verified are continuously considered correct by technicians for y times and this cluster result is different from the cluster results set by technicians, then technicians increase the weight coefficient of the corresponding clustering algorithm. This means that although the cluster result output by this clustering algorithm is not the cluster result preset by technicians, it has high accuracy. On the one hand, this can prevent clustering algorithms capable of autonomously generating cluster results from generating multiple new cluster results without restraint to interfere with technicians' assessment of the operating state of the hydrogen compressor, greatly reducing the workload of technicians. On the other hand, through a small amount of manual intervention, it can be ensured that when the hydrogen compressor actually shows a situation other than the preset cluster results of technicians, the clustering algorithm with a high weight coefficient and the ability to autonomously generate cluster results can accurately determine the operating state of the hydrogen compressor at the current monitoring moment.

[0132] In the hydrogen compressor monitoring method of the present invention, the operating state of the hydrogen compressor at the current monitoring moment can finally be accurately determined. The operating state of the hydrogen compressor not only includes the working condition of the hydrogen compressor at the current monitoring moment but also includes the prediction of the working condition of the current hydrogen compressor in a period of time in the future. Therefore, the number of preset cluster results by technicians can be more. For example, among multiple cluster results, the determination of the working condition of the hydrogen compressor at the current monitoring moment is the same, but the prediction of the working condition of the current hydrogen compressor in a period of time in the future is different. This is directly different from the prior art where only the working condition of the hydrogen compressor at the current monitoring moment can be determined by using a single thermodynamics parameter, reducing the risk of misjudgment in the prior art that may be caused by technicians presetting too many cluster results.

[0133] As Figures 2a to 2c 、 Figures 3a to 3c 、 Figures 4a to 4c shown, the abscissa is the serial number of the data segment, which is the corresponding clustering result obtained by technicians using the prior art according to different single thermodynamics parameters. Among them Figures 2a to 2c is the corresponding clustering result obtained by using the DBSCAN clustering algorithm according to different thermodynamics parameters, Figure 2a is the outlet pressure clustering result obtained based on the outlet pressure: the data during the stable period is marked in green, purple is the abnormal working condition, and a small number of normal working conditions are misidentified; Figure 2bThe clustering result of the secondary exhaust temperature obtained based on the secondary exhaust temperature: The data during the stable period is marked in purple, green for the step interval, and yellow for abnormal working conditions. It can be seen that some mutation points cannot be recognized and are mixed in the same type of working conditions; Figure 2c The clustering result of the cooler temperature obtained based on the cooler temperature. The data during the stable period is marked in purple, and yellow is identified as abnormal working conditions. The abnormal working conditions in the purple interval are not effectively identified, and yellow also belongs to misidentification, with a very high error rate. Figures 3a to 3c The corresponding clustering results obtained using the OPTICS clustering algorithm according to different thermodynamics parameters, Figure 3a The clustering result of the outlet pressure obtained based on the outlet pressure: Among them, cyan, green, and most purple are correctly identified as normal working conditions, and yellow is correctly identified as the deviation working condition, but the severely abnormal working condition is identified as the purple working condition, with misidentification here; Figure 3b The clustering result of the secondary exhaust temperature obtained based on the secondary exhaust temperature: Among them, the normal working conditions are identified as green and yellow, and the working conditions in the abnormal interval are marked as green and purple, and there are also a large number of misidentifications; Figure 3c The clustering result of the cooler temperature obtained based on the cooler temperature: The algorithm generally marks the abnormal working conditions as green and blue, and other colors are normal working conditions. However, the green mark representing abnormality still appears in some normal working conditions, and the accuracy of this algorithm is still insufficient. Figures 4a to 4c The corresponding clustering results obtained using the GMM clustering algorithm according to different thermodynamics parameters, Figure 4a The clustering result of the outlet pressure obtained based on the outlet pressure: The normal working conditions are generally marked as yellow, purple, and green, with a small amount of cyan. The abnormal interval is marked as green in large quantities and a small amount is green. However, green and cyan are distributed in both the normal and abnormal intervals, and the accuracy needs to be improved; Figure 4b The clustering result of the secondary exhaust temperature obtained based on the secondary exhaust temperature: The secondary exhaust temperature is only roughly identified as blue, yellow, and a small amount of purple, indicating that the boundary between abnormal and normal working conditions cannot be distinguished; Figure 4c The clustering result of the cooler temperature obtained based on the cooler temperature. The abnormal interval is effectively identified as the purple interval, but the green mark and the blue mark are both marked in the two working conditions, resulting in misidentification. It can be seen that the existing technology can recognize the deviation of the hydrogen compressor outlet pressure at the current monitoring moment, but cannot distinguish the degree; The abnormal working conditions in the steep rising edge interval of the secondary exhaust temperature are not recognized enough; The recognition accuracy of the cooler temperature is not high, and the variable temperature data is not effectively recognized; Different thermodynamics parameters are independent of each other.

[0134] As Figures 5a to 5c shown, the abscissa is the serial number of the data segment, which is the corresponding clustering result obtained by the monitoring method of the present invention by comprehensively considering multiple thermodynamics parameters, Figure 5aThe clustering result of the outlet pressure obtained based on the outlet pressure: for abnormal operating conditions, they are identified and marked as blue, purple, and black according to different deviation degrees, and normal operating conditions are identified as yellow and cyan; Figure 5b Based on the clustering results of the secondary exhaust temperature and the cooler temperature: the abnormal intervals are effectively identified as purple and black. Among them, the clustering result of the outlet pressure is that the present invention effectively identifies the abnormalities of the secondary exhaust temperature and the cooler temperature to automatically distinguish different pressure deviation levels. It can be seen that the abnormal rising interval of the secondary exhaust temperature, which can only be particularly reflected by the associated deviation change of the cooler in the present invention; Figure 5c Based on the clustering result of the cooler temperature obtained from the cooler temperature, the clustering result of the cooler temperature is obtained by the present invention by correlating the thermodynamics parameters such as the secondary exhaust temperature, the cooler temperature, and the outlet pressure with the deviation level, and effectively identifies the variable temperature data.

[0135] For the 6 cluster results preset by those skilled in the art, using the hydrogen compressor detection method of the present invention, compared with the GMM clustering algorithm that does not have the ability to independently generate cluster results in the prior art, a hydrogen compressor is continuously monitored for 133 hours, and the interval between adjacent monitoring times is 5 minutes. After manual verification, the accuracy rate of the monitoring method of the present invention is as high as 95.24%, while that of the prior art is only 66.73%; and during this period, the monitoring method of the present invention also generated a new cluster result as the operating state of the hydrogen compressor at the corresponding monitoring time. After verification by those skilled in the art, this operating state is correct and different from the 6 cluster results preset by those skilled in the art.

[0136] The technologies, shapes, and structures not described in detail in the present invention are all well-known technologies. It should also be noted that the above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. The components or steps in the embodiments of the present invention can be decomposed and / or recombined, and these decompositions and / or recombinations should be regarded as equivalent solutions of the present application and should all fall within the protection scope of the present invention.

Claims

1. A hydrogen compressor monitoring method based on a clustering matrix and a dynamic sliding window, characterized in that It includes the following steps: S1. For any thermodynamic parameter of the hydrogen compressor: Obtain the thermodynamic data of the current sliding window according to the window parameters of the current sliding window; The sliding windows are arranged in chronological order; S2. Calculate the statistical data of the current sliding window based on the thermodynamic data of the current sliding window; The window parameters, thermodynamic data, and statistical data of the current sliding window together constitute the current sliding window information; S3. According to the sliding window information of each thermodynamic parameter between the current monitoring moment and the previous monitoring moment, each clustering algorithm outputs the cluster result corresponding to the current monitoring moment; At the same time, calculate the weight coefficients of each clustering algorithm at the current monitoring moment; S4. Accumulate the weight coefficients of the clustering algorithms corresponding to the same cluster result as the score of the corresponding cluster result, and then take the cluster result with the highest score as the operating state of the hydrogen compressor at the current monitoring moment; After S2, it also includes S2': S2'. Calculate the window parameters of the next sliding window according to the statistical data and window parameters of the current sliding window; The following content is also included in S2': Record the current sliding window The window length of is, where t is a positive integer. The previous sliding window of the current sliding window is denoted as . Then the window length of the next sliding window is : ， ， ， ， ， Among them, represents the window length adjustment amount of the current sliding window ; represents the window length adjustment amount of the next sliding window ; is the first adjustment coefficient; is the second adjustment coefficient; represents the weight coefficient of the i-th clustering algorithm. A total of n clustering algorithms are adopted, that is, 1 ≤ i ≤ n; represents the sliding window intra-cluster entropy in the i-th clustering algorithm; represents starting from the sliding window , including the sliding window , the historical maximum entropy value of the previous m sliding windows. m is a positive integer and m ≤ t; represents the sliding window mean vector of, the modulus length of is the difference between the sliding window and the average value of the thermodynamics data of the sliding window . The direction of points from the average value of the thermodynamics data of the sliding window to the average value of the thermodynamics data of the sliding window ; represents 2-norm of; is the maximum value function; represents the sliding window intra-cluster entropy in the i-th clustering algorithm, represents the sliding window intra-cluster entropy in the i-th clustering algorithm; represents starting from the sliding window , including the sliding window , the intra-cluster entropy obtained in the corresponding clustering algorithm during the process of i ranging from 1 to n for the previous m sliding windows; represents covariance; represents variance; represents the time constant of the hydrogen compressor system; represents the current sliding window sampling period of the thermodynamics data in, which is determined by the acquisition frequency of the corresponding sensor; represents the current sliding window maximum allowable drift rate of the thermodynamics parameters of; represents the current sliding window rated value of the thermodynamics parameters of.

2. The hydrogen compressor monitoring method based on a clustering matrix and a dynamic sliding window according to claim 1, characterized in that: When monitoring the current hydrogen compressor for the first time, or during the first operation after replacing more than 1 key part of the current hydrogen compressor, for any thermodynamic parameter of the current hydrogen compressor, a new monitoring cycle has started, and the window parameters of the first sliding window in the new monitoring cycle are initial values; The window parameters include the window length and the sliding step; There are several thermodynamic data arranged in time of only one thermodynamic parameter in each sliding window.

3. A hydrogen compressor monitoring method based on a clustering matrix and a dynamic sliding window according to claim 1, characterized in that: The next sliding window The sliding step size of is as follows: = , where γ is the first constant parameter.

4. A hydrogen compressor monitoring method based on a clustering matrix and a dynamic sliding window according to claim 1, characterized in that, The next sliding window The sliding step size of is as follows: ， ， ； Among them, represents the overlapping ratio of the next sliding window ; represents the overlapping ratio of the current sliding window ; represents the attenuation coefficient of the next sliding window ; represents the target overlapping ratio; represents the total number of sliding windows that need to be transitioned currently; and are both known quantities determined by the operating state of the hydrogen compressor at the previous monitoring moment; represents the remaining number of sliding windows that need to be transitioned, including the next sliding window ; , represents the remaining number of sliding windows that need to be transitioned, including the current sliding window .

5. A hydrogen compressor monitoring method based on a clustering matrix and a dynamic sliding window according to claim 1, characterized in that The following content is also included in S3: Adopt n different clustering algorithms. Among these n clustering algorithms, some clustering algorithms do not have the ability to generate cluster results independently, and some clustering algorithms have the ability to generate cluster results independently; Send the sliding window information of each thermodynamic parameter between the current monitoring moment and the previous monitoring moment into each clustering algorithm respectively, and each clustering algorithm outputs the cluster result corresponding to the hydrogen compressor at the current monitoring moment; Calculate the weight coefficients of each clustering algorithm at the current monitoring moment simultaneously , 1 ≤ i ≤ n and i is a positive integer: If the current monitoring moment is the first monitoring moment within the current monitoring period, then obtain the weight coefficients of each clustering algorithm at the current monitoring moment based on solving the multi-objective optimization function; if the current monitoring moment is not the first monitoring moment within the current monitoring period, then obtain the weight coefficients of each clustering algorithm at the current monitoring moment based on the weight coefficients of each clustering algorithm at the previous monitoring moment.

6. The hydrogen compressor monitoring method based on the clustering matrix and the dynamic sliding window according to claim 5, wherein Based on solving the multi-objective optimization function to obtain the weight coefficients of each clustering algorithm at the current monitoring moment, the following content is also included: Construct the multi-objective optimization function F and solve it to obtain the values of the weight coefficients of each clustering algorithm at the current monitoring moment: ， ， The constraints are as follows: , Among them, is the second constant parameter; represents the weight coefficient corresponding to the i-th clustering algorithm; represents the sum of squared errors of the i-th clustering algorithm; represents the first weight matrix, which is a column vector composed of weight coefficients; represents the second weight matrix, which is the first weight matrix transposed, and is a row vector; represents the first weight matrix and the correlation matrix quadratic form; is the regularization parameter.

7. A hydrogen compressor monitoring method based on a clustering matrix and a dynamic sliding window according to claim 5, characterized in that, Based on the weight coefficients of each clustering algorithm at the previous monitoring moment, obtain the weight coefficients of each clustering algorithm at the current monitoring moment, and the following content is also included: Denote the current monitoring time as , the previous monitoring time as , z≥2 and z is a positive integer, then the weight coefficient of the i-th clustering algorithm at the current monitoring time is: ​ ， Among them, represents the weight coefficient of the i-th clustering algorithm at the previous monitoring moment; represents the weight coefficient of the j-th clustering algorithm at the previous monitoring moment; 1 ≤ j ≤ n and j is a positive integer; represents the learning rate; represents the global distribution at the previous monitoring moment; represents the distribution of the i-th clustering algorithm at the previous monitoring moment; represents the distribution of the j-th clustering algorithm at the previous monitoring moment; ) represents the JS divergence; represents the distribution and the global distribution of the JS divergence; represents the distribution and the global distribution of the JS divergence.

8. A hydrogen compressor monitoring method based on a clustering matrix and a dynamic sliding window according to claim 1, characterized in that After S4, it also includes S5: S5. If the cluster result obtained by the clustering algorithm with the ability to generate cluster results independently does not become the cluster result with the highest score, then record the cluster result obtained by the clustering algorithm with the ability to generate cluster results independently as a set to be verified. After the technician regularly checks and analyzes the set to be verified, if the cluster result obtained by a certain clustering algorithm in the set to be verified is considered correct by the technician for y consecutive times and is different from the cluster result set by the technician, then the technician increases the weight coefficient of the corresponding clustering algorithm.

Images