Real-time fault early warning method and system for labyrinth piston compressor

Abstract

The invention belongs to the field of fault early warning, particularly relates to a real-time fault early warning method and system for a labyrinth piston compressor, and aims to solve the technical problems that an existing deep learning model is insufficient in sensitivity to transient faults, and hyper-parameter optimization is prone to falling into local optimum. The early warning method comprises the following steps: S1, acquiring multi-source time sequence monitoring data, and performing normalization processing to form an input feature vector sequence; s2, constructing a time sequence feature extraction network; s3, optimizing the parameters of the network by using a partition collaborative zebra optimization algorithm; and S4, using the optimized time sequence feature extraction network to process the monitoring data in real time, outputting a fault probability sequence, and generating an early warning signal when a continuous value in the sequence is higher than an early warning threshold within a specified time length. The balance between global search and local search is realized, local optimum can be jumped out, and full optimization of parameters is ensured.

Description

Technical Field

[0001] This invention belongs to the field of fault early warning, and in particular relates to a real-time fault early warning method and system for labyrinth piston compressors. Background Technology

[0002] Labyrinth piston compressors are reciprocating compressors. To ensure equipment safety and prevent accidents, multiple types of sensors are typically deployed to monitor their operating status in real time, acquiring multi-source time-series data including vibration, piston rod displacement, cylinder pressure, and exhaust temperature. Existing fault prediction methods largely rely on deep learning models (such as gated recurrent units (GRUs)) to process the time-series data.

[0003] However, the gating mechanism of conventional GRU networks is relatively general and it is difficult to focus on the transient impact features caused by early faults that are submerged in strong background noise. Furthermore, when fusing multi-source information, features from different channels are usually simply spliced ​​or added together, resulting in insufficient feature extraction and insufficient sensitivity to early weak faults.

[0004] Furthermore, the performance of deep learning models is highly dependent on the optimized selection of their hyperparameters and network structure parameters. Traditional swarm intelligence algorithms generally suffer from the problem of effectively balancing global search and local search capabilities during the optimization process. In the early stages of iteration, excessively strong search capabilities may cause the population to quickly fall into local optima; while in the later stages of iteration, insufficient search capabilities make it difficult to escape local extrema, resulting in limited convergence accuracy and speed, thus restricting the accuracy and reliability achievable by fault warning models. Summary of the Invention

[0005] This invention provides a real-time fault early warning method and system for labyrinth piston compressors to solve the technical problems of insufficient sensitivity of existing deep learning models to transient faults and easy getting trapped in local optima during hyperparameter optimization.

[0006] The first invention provides a real-time fault early warning method for a labyrinth piston compressor, comprising the following steps: S1 acquires multi-source time-series monitoring data of compressor vibration signal, piston rod center offset, cylinder pressure and exhaust temperature during compressor operation, and forms an input feature vector sequence after normalization processing; S2, construct a temporal feature extraction network, which includes cascaded first and second gating. The first gating calculates the phase perception threshold using the instantaneous phase of the fundamental frequency component of the in-cylinder pressure. When the kurtosis value of the input feature vector within a set time window exceeds the phase perception threshold, it is passed to the second gating. The second gating is calculated using the current input feature vector, the hidden state at the previous time step, and the nonlinear cross-channel influence matrix. The nonlinear cross-channel influence matrix is ​​composed of the maximum information coefficient within the sliding time window and is weighted by the spectral entropy of the current hidden state vector to its contribution to the state update. S3 utilizes the partitioned collaborative zebra optimization algorithm to optimize network parameters, dividing the population into an exploratory subpopulation and a development subpopulation: the exploratory subpopulation adopts Levy flight, with the step size determined by the crowding factor; the development subpopulation moves towards the gradient projection optimum, and when the fitness of an individual stagnates for several consecutive generations, an opposition-based reinitialization strategy is adopted to reset the individual to the central mirror position relative to the current global optimum. S4 uses the optimized time-series feature extraction network to process monitoring data in real time and outputs a fault probability sequence. When the continuous value in the sequence is higher than the warning threshold within a specified time length, a warning signal is generated.

[0007] Furthermore, in S1, after normalizing the multi-source time-series monitoring data, it is divided into data segments of fixed length to obtain input samples. The input samples are structured input feature vector sequences. Each input feature vector corresponds to a sampling time and includes the vibration signal, piston rod center offset, cylinder pressure, and exhaust temperature at that sampling time.

[0008] Furthermore, in S2, the formula for calculating the phase sensing threshold is: ; in, Let be the phase sensing threshold at time t. The preset baseline kurtosis threshold, The instantaneous phase of the fundamental frequency component is obtained by performing a Hilbert transform on the cylinder pressure signal. The modulation coefficient takes values ​​in the interval (0, 1).

[0009] Furthermore, the interval where the instantaneous phase is in the compression and power stroke is designated as the high-sensitivity zone, and a lower phase perception threshold is set accordingly; the remaining phase intervals are designated as the low-sensitivity zone, and a higher threshold is set accordingly.

[0010] Furthermore, in S2, within each sliding time window, the maximum information coefficient value between each pair of the four monitoring data channels—vibration signal, piston rod center offset, cylinder pressure, and exhaust temperature—is calculated. The calculated six maximum information coefficient values ​​and the value 1 on the diagonal are used to construct a 4×4 influence matrix; Calculate the spectral entropy of the current hidden state vector, and use the spectral entropy as a weighting factor to weight the nonlinear cross-channel influence matrix in the hidden state update calculation.

[0011] Furthermore, in S3, the optimization variables for the temporal feature extraction network include the learning rate and the number of hidden layer nodes.

[0012] Furthermore, in S3, for any individual in the exploration subpopulation, the five nearest neighbor individuals with Euclidean distance are found in the parameter optimization space; Calculate the average Euclidean distance between this individual and its five neighboring individuals. ; The crowding factor C is calculated using the following formula: ; in It is a very small positive number used to prevent the denominator from being zero; The final movement step size is obtained by multiplying the base step size generated by the Levi flight strategy by the crowding factor C.

[0013] Furthermore, in S3, the process of determining the optimal point of gradient projection includes the following steps: Select the three individuals with the highest fitness in the current population; Obtain the parameter space positions of these three individuals in the current generation t and the previous generation t-1. and ; Calculate the average position change vector ; Gradient projection optimal point The position of the current global best individual The result is obtained by extrapolation, and the calculation formula is: ;in, It is an adjustable extrapolation step size factor used to control the extent of extension in the prediction direction.

[0014] Furthermore, in S3, an opposition-based reinitialization strategy is adopted to reset individuals to a central mirror position relative to the current global optimum, including the following steps: A fitness stagnation counter is set for each individual in the development subpopulation. When the fitness of an individual does not improve for 5 consecutive generations, a re-initialization process is triggered. Obtain the current position vector of the individual in the parameter space. and the position vector of the current global optimal solution Calculate the center mirror position vector relative to the current global optimum. The calculation formula is as follows: ; The individual's position is updated to the calculated center mirror position vector. .

[0015] The second invention provides a real-time fault warning system for a labyrinth piston compressor, including a memory and a processor. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned real-time fault warning method for a labyrinth piston compressor is implemented.

[0016] The beneficial effects are as follows: By calculating the phase perception threshold through the instantaneous phase of the fundamental frequency component of the cylinder pressure, the first gating can filter out transient impact features related to the equipment's operating cycle based on the phase perception threshold, thus suppressing background noise interference. The second gating, by constructing a cross-channel influence matrix based on the maximum information coefficient and using spectral entropy for contribution weighting, achieves information fusion and improves the ability to extract early, weak fault features. Simultaneously, the partitioned collaborative zebra optimization algorithm, through exploring and developing the division of labor and cooperation among subpopulations, combined with gradient extrapolation-based optimization guidance and an adversarial re-initialization strategy for stagnant individuals, achieves a balance between global and local search, enabling it to escape local optima and ensuring sufficient parameter optimization. Attached Figure Description

[0017] Figure 1 A flowchart for a real-time fault early warning method for a labyrinth piston compressor; Figure 2 This is a schematic diagram illustrating the dynamic relationship between instantaneous phase, phase sensing threshold, and input feature kurtosis. Figure 3 A schematic diagram illustrating the division between exploring and developing subpopulations; Figure 4 A schematic diagram for calculating individual crowding in a subpopulation; Figure 5 A schematic diagram illustrating the path for a subpopulation to move towards the optimal point of gradient projection; Figure 6 This is a block diagram of a real-time fault warning system for a labyrinth piston compressor. Detailed Implementation

[0018] An embodiment of the real-time fault early warning method for labyrinth piston compressors provided by this invention: like Figure 1 As shown, a real-time fault early warning method for a labyrinth piston compressor includes the following steps: S1 acquires multi-source time-series monitoring data of compressor operation, including vibration signal, piston rod center offset, cylinder pressure, and exhaust temperature, and forms an input feature vector sequence after normalization.

[0019] Install sensors in key parts of the compressor: install an acceleration sensor on the compressor body to collect vibration signals, install an eddy current sensor near the piston rod to monitor the piston rod center offset, install a pressure sensor on the cylinder head to obtain the cylinder pressure, and install a thermocouple in the exhaust pipe to measure the exhaust temperature.

[0020] Four signals are simultaneously acquired using a data acquisition card to obtain multi-source time-series monitoring data. This data is then normalized and divided into fixed-length data segments to obtain input samples. For example, each data segment contains 1024 sampling points. The input samples are structured sequences of input feature vectors, each corresponding to a sampling time and containing the vibration signal, piston rod center offset, cylinder pressure, and exhaust temperature at that sampling time.

[0021] S2 constructs a temporal feature extraction network, which includes cascaded first and second gating mechanisms. The first gating mechanism calculates the phase perception threshold using the instantaneous phase of the fundamental frequency component of the in-cylinder pressure. When the kurtosis value of the input feature vector within a set time window exceeds the phase perception threshold, it is passed to the second gating mechanism. The second gating mechanism is calculated using the current input feature vector, the hidden state at the previous time step, and the nonlinear cross-channel influence matrix. The nonlinear cross-channel influence matrix is ​​composed of the maximum information coefficient within the sliding time window and is weighted by the spectral entropy of the current hidden state vector to increase its contribution to the state update.

[0022] The basic architecture of the temporal feature extraction network is a gated recurrent unit (GRU), but its internal update and reset gates are replaced by a dual-gated unit: the first gating is used to filter out background noise and obtain impact features. Specifically, the compressor frequency is obtained by performing a Fourier transform on the in-cylinder pressure signal, and the compressor frequency is used as the fundamental frequency; then, the instantaneous phase of the fundamental frequency is calculated through a Hilbert transform. According to the compressor's working stroke, the interval where the instantaneous phase is in the compression and power strokes is designated as the high-sensitivity region, and a lower phase sensing threshold (e.g., 2.5) is set accordingly; the remaining phase intervals are designated as the low-sensitivity region, and a higher threshold (e.g., 5.0) is set accordingly. Figure 2 As shown.

[0023] When the kurtosis value of the input feature vector within a set time window (e.g., a time interval consisting of 10-50 consecutive sampling points) is greater than the phase perception threshold corresponding to the current phase, the first gating opens, passing the input feature vector to the second gating; otherwise, a zero vector is output. The second gating is responsible for information fusion. Within a sliding time window of 200 data points, it calculates the maximum information coefficients between each pair of the four channels (vibration, displacement, pressure, and temperature), forming a 4×4 influence matrix. This influence matrix is ​​multiplied by the input feature vector to achieve information interaction. Simultaneously, it calculates the spectral entropy of the hidden state vector from the previous moment, normalizes it, and uses it as a weight to adjust the contribution of the influence matrix to the hidden state update. A higher spectral entropy indicates a more unstable state, and the greater the contribution weight of the influence matrix. The second gating integrates the current input feature vector, the hidden state from the previous moment, and the weighted cross-channel influence information to calculate and generate a new hidden state.

[0024] In a more specific embodiment, the formula for calculating the phase sensing threshold is:

[0025] in, Let be the phase sensing threshold at time t. The preset baseline kurtosis threshold is the basic reference value for the phase-sensing threshold, reflecting the basic criteria for judging kurtosis anomalies. The instantaneous phase of the fundamental frequency component is obtained by performing a Hilbert transform on the cylinder pressure signal. The modulation coefficient takes values ​​in the interval (0, 1).

[0026] For example, when monitoring a normally operating compressor, a baseline kurtosis threshold can be preset. The value is 4.5, which is an empirical value slightly higher than the average signal kurtosis under normal operating conditions. The modulation coefficient is also set. The value is 0.2. When the piston reaches near the top dead center of the compression stroke, the cylinder pressure reaches its peak value, at which point the instantaneous phase of the fundamental frequency component is... near , The value approaches 1, and the calculated phase sensing threshold is... It rose to 5.4.

[0027] At the moments when valves are most prone to leakage (such as during compression and power strokes), lowering the phase sensing threshold can enhance the system's sensitivity to real fault impact signals, ensuring that subtle fault characteristics are not missed. Conversely, when the piston is in the middle position of the intake or exhaust stroke, the instantaneous phase... Possibly close , The value approaches -1, at which point the phase sensing threshold... This will be reduced to 3.6. In these non-critical phases, a higher phase sensing threshold is sufficient to determine the presence of abnormal shocks. A phase sensing threshold that changes synchronously with the equipment's operating cycle reduces false alarms caused by fluctuations in normal operating conditions, while simultaneously improving the ability to detect early, subtle fault characteristics at critical moments.

[0028] In a more specific embodiment, in S2, within each sliding time window, the maximum information coefficient value between each pair of the four monitoring data channels—vibration signal, piston rod center offset, cylinder pressure, and exhaust temperature—is calculated. The calculated six maximum information coefficient values ​​and the value 1 on the diagonal are used to construct a 4×4 influence matrix; Calculate the spectral entropy of the current hidden state vector, and use the spectral entropy as a weighting factor to weight the nonlinear cross-channel influence matrix in the hidden state update calculation.

[0029] For example, a sliding time window of 256 data points is set. Assume that at a certain moment, vibration, piston rod displacement, cylinder pressure, and exhaust temperature data of 256 points are collected. The calculated maximum information coefficient between the vibration signal and the cylinder pressure signal is 0.85, indicating a strong nonlinear coupling relationship between the two, possibly due to vibration and pressure fluctuations caused by valve impact. Simultaneously, the calculated maximum information coefficient between vibration and exhaust temperature is 0.2, indicating a weak correlation between the two.

[0030] The maximum information coefficient values ​​of all six cross-channels are filled into the off-diagonal positions of a 4×4 matrix, with the diagonal positions set to 1, forming the influence matrix for the current moment. The model calculates that the spectral entropy value of its internal hidden state vector is 0.9, indicating that the model's judgment of the current equipment state has high uncertainty, possibly indicating that it is in the early stages of state transition or fault evolution. Therefore, 0.9 is used as a weight and multiplied by the entire nonlinear cross-channel influence matrix. In subsequent model state updates, information from strongly correlated channels such as vibration and cylinder pressure carries a greater weight in their respective update calculations, thus helping the model to more quickly utilize multi-source information to identify fault modes.

[0031] S3 optimizes the network parameters using the partitioned collaborative zebra optimization algorithm, dividing the population into an exploratory subpopulation and a development subpopulation: the exploratory subpopulation uses Levy flight, with the step size determined by the crowding factor; the development subpopulation moves towards the gradient projection optimum. When an individual's fitness stagnates for several consecutive generations, an opposition-based reinitialization strategy is adopted to reset the individual to the central mirror position relative to the current global optimum.

[0032] In one embodiment, hyperparameters such as the learning rate and number of hidden layer nodes of the temporal feature extraction network are used as optimization variables to initialize a zebra population of 50 individuals. The alert accuracy of the temporal feature extraction network on the validation set is used as the fitness metric. The top 15 individuals in terms of fitness are assigned to the development subpopulation, and the remaining 35 individuals constitute the exploration subpopulation. Figure 3 As shown.

[0033] The subpopulation is explored using a Lévy flight strategy to update positions, with the step size scaled by a crowding factor. Crowding is determined by the average Euclidean distance of the individual to its five nearest neighbors in the parameter space: the greater the distance, the lower the crowding, and the larger the crowding factor. Figure 4 As shown, this encourages individuals to conduct large-scale searches in sparse areas.

[0034] Developing a subpopulation involves moving towards a gradient projection optimum. This optimum is obtained by selecting the three individuals with the highest current fitness, analyzing their positions and fitness trends over the past five generations, fitting gradient directions to each, and then performing linear extrapolation on these gradient directions. Figure 5 As shown, this is to accelerate local fine-grained search.

[0035] During the iteration process, if the fitness of an individual in the development subpopulation does not improve for 10 consecutive generations, it is considered to have stagnated. At this time, an opposition-based reinitialization strategy is adopted. Based on its current position, the boundary of the search space, and the position of the current global optimum, its central mirror position in the search space is calculated, and the individual is reset to this new position to force it out of the local optimum and restore its search vitality.

[0036] In a more specific embodiment, in S3, for any individual in the exploration subpopulation, the five nearest neighbor individuals with Euclidean distance are found in the parameter optimization space; Calculate the average Euclidean distance between this individual and its five neighboring individuals. ; The crowding factor C is calculated using the following formula:

[0037] in It is a very small positive number used to prevent the denominator from being zero; The final movement step size is obtained by multiplying the base step size generated by the Levi flight strategy by the crowding factor C.

[0038] In the iterative process of finding the optimal hyperparameters for the temporal feature extraction network, assume that the parameter coordinates of an individual A in the exploration subpopulation are [10.5, 3.2, 0.8]. By searching its surroundings, the Euclidean distances of its five nearest neighbors are found to be 0.11, 0.13, 0.14, 0.16, and 0.18, respectively. These neighbors are very close to individual A, indicating that individual A is in a very crowded region, and its average distance... The value is 0.144. The calculated crowding factor C is approximately 6.94. If the Levy flight strategy generates a base step size of 0.5 for this individual, then its final movement step size will be amplified to 3.47. The larger step size will cause individual A to make a large jump, leaving the crowded region that may lead to premature convergence, and to search for other unknown areas in the parameter space.

[0039] Conversely, assuming another individual B is in a sparse region, the average distance of its nearest neighbor is... If the value is 5.2, then the crowding factor C is only 0.19, which will cause the movement step size to shrink to 0.095.

[0040] In a more specific embodiment, the process of determining the optimal point of gradient projection includes the following steps: Select the three individuals with the highest fitness in the current population; Obtain the parameter space positions of these three individuals in the current generation t and the previous generation t-1. and ; Calculate the average position change vector ; Gradient projection optimal point The position of the current global best individual The result is obtained by extrapolation, and the calculation formula is: ;in, It is an adjustable extrapolation step size factor used to control the extent of extension in the prediction direction.

[0041] Suppose that during the 10th generation of optimization, the three individuals with the highest fitness were positioned at [2.1, 5.4], [2.3, 5.2], and [2.2, 5.5] in the 9th generation, and moved to [2.15, 5.48], [2.34, 5.25], and [2.26, 5.53] in the 10th generation, respectively. The position change vectors of the three elite individuals are [0.05, 0.08], [0.04, 0.05], and [0.06, 0.03].

[0042] First, calculate the average position change vector. That is, the direction of the average gradient: This indicates the co-evolutionary trend of the optimal group of individuals in the current population.

[0043] Next, find the globally optimal individual in the entire population and its position. The value is [2.35, 5.26]. Based on this optimal position, extrapolation is performed along the previously calculated average gradient direction, and the value is set to... Given a value of 1.2, calculate the optimal point of the gradient projection. The newly calculated point is [2.41, 5.324]. As a gravitational center for other individuals in the development subpopulation, it guides them to concentrate their search towards areas predicted to be more promising.

[0044] In a more specific embodiment, in S3, an opposition-based reinitialization strategy is adopted to reset the individual to a central mirror position relative to the current global optimum, including the following steps: A fitness stagnation counter is set for each individual in the development subpopulation. When the fitness of an individual does not improve for 5 consecutive generations, a re-initialization process is triggered. Obtain the current position vector of the individual in the parameter space. and the position vector of the current global optimal solution Calculate the center mirror position vector relative to the current global optimum. The calculation formula is as follows: ; The individual's position is updated to the calculated center mirror position vector. .

[0045] Suppose that for one individual, the diagnostic accuracy remains at 98.1% for five consecutive generations (from generation 20 to 24), without any improvement. At this point, its stall counter reaches 5, triggering a reinitialization process. The current parameter position vector of this individual... The value is [15.2, 8.8, -3.1]. The known global optimal solution's position vector is within the entire population. The values ​​are [12.0, 7.5, -4.0], with an accuracy of 99.2%.

[0046] To allow this stagnant individual to escape its current local optimum, we compute its counterpart with respect to the global optimum. Its central mirror image position vector... The value is calculated using the formula as [8.8, 6.2, -4.9]. The position of the stagnant individual will be updated to the newly calculated position [8.8, 6.2, -4.9]. This is equivalent to moving the individual from its current position to the other side of the parameter space symmetrical to the global optimum, increasing the possibility of finding a better solution.

[0047] S4 uses the optimized time-series feature extraction network to process monitoring data in real time and outputs a fault probability sequence. When the continuous value in the sequence is higher than the warning threshold within a specified time length, a warning signal is generated.

[0048] The optimized temporal feature extraction network is deployed in the online monitoring system to receive data streams from four sensors in real time. The network processes data continuously using a sliding window approach, and its output layer outputs a value between 0 and 1 via a sigmoid function, representing the probability of compressor failure at the current moment. A warning threshold of 0.8 is set for a time interval of 3 seconds. This means that if the network outputs a failure probability value greater than 0.8 for three consecutive seconds, an alarm is automatically triggered, sending a warning message to the operator indicating a potential early-stage equipment failure requiring inspection.

[0049] For example, an online monitoring system outputs a failure probability value every 0.1 seconds regarding whether the piston rings of a device are worn. The warning threshold is set to 0.8, meaning an alarm is only considered when the time-series feature extraction network is highly certain of the existence of a fault; the time interval is set to 3 seconds, implying that continuous evidence is required.

[0050] At a certain point in time, the output probability sequence is 0.82, 0.85, 0.91, 0.79, 0.88, ... Although the first three values ​​and the fifth value exceed the threshold of 0.8, the fourth value of 0.79 falls below the threshold, the continuous timer is reset, and no warning is issued.

[0051] However, during another time period, the output probability sequence was 0.81, 0.83, 0.86, ..., and for the next 3 seconds, the output values ​​for 30 consecutive seconds were all above 0.8. When the 30th value was 0.92, the condition of being above 0.8 for 3 consecutive seconds was met, and a clear piston ring wear warning signal was sent to the central control room, prompting maintenance personnel to pay attention.

[0052] An embodiment of the real-time fault early warning system for labyrinth piston compressors provided by this invention: like Figure 6 As shown, the real-time fault warning system for the labyrinth piston compressor includes a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned real-time fault warning method for the labyrinth piston compressor is implemented.

[0053] The real-time fault warning system for labyrinth piston compressors also includes other components well known to those skilled in the art, such as communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0054] In addition, in the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

Claims

1. A real-time fault early warning method for a labyrinth piston compressor, characterized in that, Includes the following steps: S1 acquires multi-source time-series monitoring data of compressor vibration signal, piston rod center offset, cylinder pressure and exhaust temperature during compressor operation, and forms an input feature vector sequence after normalization processing; S2, construct a temporal feature extraction network, which includes cascaded first and second gating. The first gating calculates the phase perception threshold using the instantaneous phase of the fundamental frequency component of the in-cylinder pressure. When the kurtosis value of the input feature vector within a set time window exceeds the phase perception threshold, it is passed to the second gating. The second gating is calculated using the current input feature vector, the hidden state at the previous time step, and the nonlinear cross-channel influence matrix. The nonlinear cross-channel influence matrix is ​​composed of the maximum information coefficient within the sliding time window and is weighted by the spectral entropy of the current hidden state vector to its contribution to the state update. S3 utilizes the partitioned collaborative zebra optimization algorithm to optimize network parameters, dividing the population into an exploratory subpopulation and a development subpopulation: the exploratory subpopulation adopts Levy flight, with the step size determined by the crowding factor; the development subpopulation moves towards the gradient projection optimum, and when the fitness of an individual stagnates for several consecutive generations, an opposition-based reinitialization strategy is adopted to reset the individual to the central mirror position relative to the current global optimum. S4 uses the optimized time-series feature extraction network to process monitoring data in real time and outputs a fault probability sequence. When the continuous value in the sequence is higher than the warning threshold within a specified time length, a warning signal is generated.

2. The real-time fault early warning method for a labyrinth piston compressor according to claim 1, characterized in that, In S1, the multi-source time-series monitoring data is normalized and then divided into data segments of fixed length to obtain input samples. The input samples are structured input feature vector sequences. Each input feature vector corresponds to a sampling time and includes the vibration signal, piston rod center offset, cylinder pressure, and exhaust temperature at that sampling time.

3. The real-time fault early warning method for a labyrinth piston compressor according to claim 1, characterized in that, In S2, the formula for calculating the phase sensing threshold is: ; wherein is a phase-aware threshold for time t, is a preset reference kurtosis threshold, is an instantaneous phase of a fundamental component obtained by Hilbert transform of the in-cylinder pressure signal, is a modulation coefficient taking values in the interval (0, 1).

4. The real-time fault early warning method for a labyrinth piston compressor according to claim 3, characterized in that, The interval where the instantaneous phase is in the compression and power stroke is designated as the high-sensitivity zone, and a lower phase perception threshold is set accordingly; the remaining phase intervals are designated as the low-sensitivity zone, and a higher threshold is set accordingly.

5. The real-time fault early warning method for a labyrinth piston compressor according to claim 1, characterized in that, In S2, within each sliding time window, the maximum information coefficient value between each pair of the four monitoring data channels—vibration signal, piston rod center offset, cylinder pressure, and exhaust temperature—is calculated. The calculated six maximum information coefficient values ​​and the value 1 on the diagonal are used to construct a 4×4 influence matrix; Calculate the spectral entropy of the current hidden state vector, and use the spectral entropy as a weighting factor to weight the nonlinear cross-channel influence matrix in the hidden state update calculation.

6. The real time fault warning method for labyrinth piston compressor as claimed in claim 1 wherein, In S3, the optimization variables for the temporal feature extraction network include the learning rate and the number of hidden layer nodes.

7. The real time fault warning method for labyrinth piston compressor as claimed in claim 1 wherein, In S3, for any individual in the exploration subpopulation, find the 5 nearest neighbor individuals in the parameter optimization space; Calculate the average Euclidean distance between this individual and the 5 neighbor individuals ; The congestion factor C is calculated by the following equation: ; wherein is a very small positive number to prevent the denominator from being zero; The final movement step size is obtained by multiplying the base step size generated by the Levi flight strategy by the crowding factor C.

8. The real-time fault early warning method for a labyrinth piston compressor according to claim 7, characterized in that, In S3, the process of determining the optimal point of gradient projection includes the following steps: Select the three individuals with the highest fitness in the current population; obtaining the parameter space positions of the three individuals at the current generation t and the previous generation t-1 and ; Calculate the average position change vector ; Gradient projection optimal point The position of the current global best individual The result is obtained by extrapolation, and the calculation formula is: ;in, It is an adjustable extrapolation step size factor used to control the extent of extension in the prediction direction.

9. The real-time fault early warning method for a labyrinth piston compressor according to claim 8, characterized in that, In S3, an opposition-based reinitialization strategy is adopted to reset individuals to a central mirror position relative to the current global optimum, including the following steps: A fitness stagnation counter is set for each individual in the development subpopulation. When the fitness of an individual does not improve for 5 consecutive generations, a re-initialization process is triggered. Obtain the current position vector of the individual in the parameter space. and the position vector of the current global optimal solution Calculate the center mirror position vector relative to the current global optimum. The calculation formula is as follows: ; updating the position of the individual to the calculated center mirror position vector .

10. A real-time fault warning system for a labyrinth piston compressor, characterized in that It includes a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the real-time fault early warning method for the labyrinth piston compressor as described in any one of claims 1-9 is implemented.

Images