An abnormality identification system and method for an oil depot pump set

Abstract

The application relates to the field of equipment state monitoring, and particularly discloses an oil depot pump set abnormality identification system and method, which comprises a multi-type data acquisition module, an adaptive threshold adjustment module, a multi-modal deep learning fusion identification module, an abnormality trend prediction module and a correlation decision support module; the multi-type data acquisition module is used for acquiring vibration signals, physical state signals of an oil depot pump set and characteristic parameters of an operation environment and a load where the oil depot pump set is located, converting the vibration signals into two-dimensional time-varying spectrum features, converting the physical state signals into standard time sequence data, and transmitting the standard time sequence data to the multi-modal deep learning fusion identification module; the technical scheme can reduce false alarm and missed alarm probabilities, realize deep fusion of data body features of multi-type sensors, and can guide maintenance decisions according to state monitoring.

Description

Technical Field

[0001] This invention relates to the field of equipment condition monitoring, and in particular to an abnormal identification system and method for oil depot pump sets. Background Technology

[0002] As the core power equipment in oil transportation and storage systems, the safety and stability of oil depot pump sets play a decisive role in the overall safe production of the oil depot. If abnormalities such as mechanical wear, bearing failure, pump leakage, or motor malfunction occur during pump set operation and are not detected in time, it could lead to catastrophic production safety accidents. Therefore, configuring an efficient and accurate pump set anomaly identification and operation and maintenance decision-making system is crucial for oil depot safety.

[0003] Traditional methods for monitoring anomalies in oil depot pump units primarily rely on single-type physical quantity sensors (such as vibration sensors or temperature sensors alone) deployed on the pump body, and set fixed upper and lower alarm thresholds to determine whether the pump unit has malfunctioned. However, this monitoring method has significant shortcomings in practical applications.

[0004] The reason is that the judgment threshold of the oil depot pump set lacks dynamic adaptability, resulting in a high probability of false alarms and missed alarms. The actual operating status of the oil depot pump set is easily affected by various dynamic factors such as ambient temperature and humidity, and equipment load level. Under different ambient temperatures and working loads, even if the pump set is in a completely healthy state, its characteristic physical quantities (such as temperature, current, and vibration) will fluctuate normally. If a fixed alarm threshold is used, when the equipment is operating under high temperature or high load conditions, it is very easy to trigger false alarms due to normal fluctuations; while when the equipment is operating under low temperature or low load conditions, even if weak or early abnormal characteristics have been generated, its physical quantities may not exceed the fixed threshold, thus causing missed alarms.

[0005] In addition, the data volume characteristics of different types of sensors vary significantly, making it difficult to achieve deep spatiotemporal feature fusion. Different types of sensors have vastly different physical sampling frequencies (for example, vibration sensors capturing high-frequency mechanical vibrations typically require sampling frequencies in the kilohertz range, while temperature sensors monitoring bearing temperature only require sampling frequencies in the hertz range). Currently, isolated thresholding or simple linear splicing of single-sequence data collected by each sensor is commonly used, which fails to achieve spatiotemporal alignment and collaborative analysis of heterogeneous multimodal data. This makes it difficult to deeply explore the intrinsic correlations between motor operating current fluctuations, pump impedance and pressure changes, and high-frequency vibration characteristics, thus hindering the accurate characterization of the complex nonlinear evolution characteristics within the pump unit.

[0006] There are also issues such as a disconnect between condition monitoring and maintenance decision-making processes, delayed operation and maintenance response, and unreasonable allocation of maintenance resources. Existing anomaly diagnosis systems are typically limited to judging the current instantaneous state and lack the ability to predict the development trend of anomalies over a future time span. Furthermore, traditional equipment alarms only provide information on single physical quantity exceeding limits, failing to establish intelligent linkage between anomaly characteristics, evolution risks, and maintenance tasks. This forces maintenance personnel to manually formulate maintenance plans and allocate parts based on experience after receiving alarms, severely impacting equipment operation and maintenance efficiency and increasing unplanned downtime.

[0007] Therefore, there is an urgent need for an oil depot pump group anomaly identification system that can reduce the probability of false alarms and false alarms, achieve deep fusion of data volume characteristics from multiple types of sensors, and guide maintenance decisions based on condition monitoring. Summary of the Invention

[0008] This invention provides an oil depot pump group anomaly identification system that can reduce the probability of false alarms and missed alarms, achieve deep fusion of data volume features from multiple types of sensors, and guide maintenance decisions based on condition monitoring.

[0009] To solve the above-mentioned technical problems, this application provides the following technical solution:

[0010] An oil depot pump group anomaly identification system includes a multi-type data acquisition module, an adaptive threshold adjustment module, a multi-modal deep learning fusion identification module, an anomaly trend prediction module, and an association decision support module;

[0011] The multi-type data acquisition module is used to acquire vibration signals, physical state signals, and characteristic parameters of the operating environment and load of the oil depot pump group. It converts the vibration signals into two-dimensional time-varying spectrum features and the physical state signals into standard time series data, which are then transmitted to the multimodal deep learning fusion recognition module. At the same time, the characteristic parameters are transmitted to the adaptive threshold adjustment module and the abnormal trend prediction module.

[0012] The adaptive threshold adjustment module is used to perform mean clustering calculation on historical feature parameters to obtain the centroid of the working condition division, determine the environmental working condition category according to the proximity of the currently collected feature parameters to the centroid of the working condition division, match the corresponding deviation adjustment coefficient according to the environmental working condition category, and correct the benchmark abnormal threshold based on the deviation adjustment coefficient to generate an adaptive dynamic judgment threshold and transmit it to the multimodal deep learning fusion recognition module.

[0013] The multimodal deep learning fusion recognition module is used to extract the local spatiotemporal joint features of the two-dimensional time-varying spectrogram using a convolutional neural network architecture, and to extract the dynamic temporal correlation features of the standard time series data over time using a long short-term memory network architecture. The extracted local spatiotemporal joint features and dynamic temporal correlation features are concatenated into feature vectors to obtain multimodal features. A self-attention weighting mechanism is introduced to assign sensitivity weights to each physical quantity dimension of the multimodal features for feature weighting. The weighted multimodal features are numerically compared with the adaptive dynamic judgment threshold to obtain the current anomaly probability.

[0014] The abnormal trend prediction module is used to extract the time-varying evolution trend of the standard time series data within a set sliding window scale, along with the current feature parameters, into a deep recurrent hybrid network to predict the probability curve of anomalies occurring within a set future time period.

[0015] The associated decision support module is used to input the current anomaly probability and the anomaly occurrence probability curve, use series comparison logic to retrieve and match the corresponding fault safety risk level, and interact with the task scheduling database interface to generate appropriate operation and maintenance scheduling instructions.

[0016] The basic principle and beneficial effects of this invention are as follows: The multi-type data acquisition module transforms the highly sampled vibration signal into a two-dimensional time-varying spectrum feature and the low-sampled physical state signal into standard time-series data. Through the convolutional neural network architecture and long short-term memory network architecture in the multi-modal deep learning fusion recognition module, it extracts image-based local spatiotemporal joint features and one-dimensional waveform dynamic temporal correlation features, respectively. After obtaining multi-modal features through feature concatenation, the self-attention weighting mechanism is introduced. This architecture achieves spatiotemporal scale alignment of heterogeneous data at the feature layer and can automatically allocate higher weights to key sensor channels and sensitive time intervals using self-attention weight calculation. Compared to the simple data concatenation or independent analysis of existing technologies, this principle significantly improves the sensitivity and classification accuracy for identifying weak, gradual, and time-varying nonlinear fault modes such as pump rotor jamming, bearing spalling, and shaft misalignment. It also effectively suppresses environmental background noise in a single channel, improving overall anti-interference capability.

[0017] This invention utilizes an adaptive threshold adjustment module to perform mean clustering calculations based on historical feature parameters to determine the centroid of operating condition classification. It then dynamically retrieves the corresponding environmental operating condition category based on current parameters, and adaptively corrects the baseline anomaly threshold using the deviation adjustment coefficient. This mechanism breaks the limitations of traditional fixed judgment thresholds, allowing the anomaly judgment boundary to automatically and smoothly deform with diurnal temperature differences and pumping load fluctuations, forming a dynamically wrapped safety judgment threshold envelope. By appropriately relaxing the adaptive dynamic judgment threshold under high temperature and heavy load conditions, and appropriately tightening it under low temperature and light load conditions where minor faults are easily hidden, the invention significantly reduces frequent false alarms caused by normal physical fluctuations in equipment. Simultaneously, it avoids fatal missed detections caused by early weak faults being masked by normal operating condition biases under low operating conditions, significantly enhancing the system's detection robustness under variable climates and complex load conditions.

[0018] This invention utilizes the anomaly trend prediction module to predict future trends of historical sequence data and real-time feature parameters within a sliding window scale, obtaining the anomaly occurrence probability curve. The associated decision support module simultaneously inputs the current anomaly probability and the anomaly occurrence probability curve representing future degradation. Compared to traditional single-state over-limit alarms, it can accurately determine the comprehensive fault safety risk level using series comparison logic and interact with the task scheduling database interface to generate appropriate operation and maintenance scheduling instructions. This closed-loop linkage mechanism transforms the monitoring system into a predictive proactive protection system, enabling the early initiation of spare parts preparation and maintenance work order dispatch before the unit reaches its shut-down limit and when the fault is in its early stages. This significantly shortens the response time for operation and maintenance decisions, substantially reduces unplanned pump downtime, and lowers the operation and maintenance costs throughout the entire lifecycle of the oil depot assets.

[0019] At the data processing level, the data acquisition module first transforms and reduces the dimensionality of the signal at the front end, avoiding the direct input of the massive amount of raw high-frequency vibration waveform data into the main network for computation. This significantly reduces the processor's computing power consumption and processing latency during online diagnosis, ensuring the system's millisecond-level real-time online monitoring response speed at the edge. Simultaneously, the correlation decision support module uses the time-series development trend provided by the anomaly trend prediction module to collaboratively verify the current anomaly probability. For single-point probability spikes caused by non-structural transient impacts such as instantaneous power grid pulsations and pipeline cavitation, it can be detected that the predicted probability curve does not show a continuously deteriorating trend, thus achieving logic debouncing and eliminating production fluctuations caused by frequent machine switching due to occasional brief disturbances, ensuring the stable operation of the oil depot's oil transportation process.

[0020] In summary, this invention achieves the goals of reducing the probability of false alarms and missed alarms, realizing deep fusion of data volume features from multiple types of sensors, and guiding maintenance decisions based on condition monitoring.

[0021] Furthermore, the multi-type data acquisition module includes an external piezoelectric vibration sensor, an armored temperature sensor, an inlet pressure probe and an outlet pressure probe, and a three-phase current transformer for the motor circuit, all of which are uniformly connected to a bus that supports a high-precision time transmission protocol to align the time-domain phase of each data source.

[0022] Furthermore, the multi-type data acquisition module also includes a preprocessing arithmetic unit. The preprocessing arithmetic unit uses a preset low-frequency and high-frequency amplitude-limiting bandpass digital filter to extract the time-domain fluctuation component of the high-frequency vibration from the voltage signal output by the external piezoelectric vibration sensor. It also applies a mapping normalization formula to normalize the measured values ​​of the armored temperature sensor, the inlet pressure probe, the outlet pressure probe, and the three-phase current transformer of the motor circuit to a symmetrical open interval range.

[0023] Furthermore, the multimodal deep learning fusion recognition module includes a convolutional neural computation unit and a long short-term memory neural network unit; the convolutional neural computation unit performs local filtering convolution operations on the two-dimensional time-varying spectrum features based on short-time Fourier transform, and the long short-term memory neural network unit extracts the standard time series data with a self-gated layer structure, and concatenates the calculation results of the two onto a unified feature axis to complete the splicing.

[0024] Furthermore, the multimodal deep learning fusion recognition module also includes a correlation weighting processor; the correlation weighting processor converts all the spliced ​​features to the query layer, the feature association key layer and the real mapping value layer, performs two-vector dot product multiplication and implements amplitude scaling normalization, and reconstructs the response gain at different sensor representation components to amplify the feature intensity of the abnormal channel;

[0025] The adaptive dynamic judgment threshold is obtained by multiplying the preset value of the baseline operating state by the environmental adjustment value and superimposing it. The environmental adjustment value is within the preset maximum bidirectional deviation amplitude ratio range. The adaptive threshold adjustment module performs parameter relearning of the clustering model for the environmental condition category at the end of each day and re-adjusts the deviation adjustment coefficient.

[0026] Furthermore, it also includes a data interconnection platform composed of distributed edge control nodes distributed in the pump room and cloud storage resources; the distributed edge control nodes adopt a lightweight transmission protocol with transmission loss prevention verification, aggregate the processed packet signals into timed compressed data frames and send them to the cloud, and use a distributed write-only time-series database to achieve physical archiving according to the data generation time scale.

[0027] Furthermore, the data interconnection platform includes an adaptive frequency conversion acquisition controller; the adaptive frequency conversion acquisition controller compares the current anomaly probability and the anomaly occurrence probability curve with a safety standard line; if it exceeds the safety standard line, it sends an acceleration control command to the distributed edge control node to increase the sampling rate of the external piezoelectric vibration sensor and initiates a local high-speed anti-loss physical dump mechanism; if it does not exceed the safety standard line, it automatically reduces the data frame frequency to switch to a low-power intermittent scanning state.

[0028] Furthermore, it also includes a progressive offset automatic correction unit; the progressive offset automatic correction unit monitors the cumulative prediction error generated by the current abnormal probability. If the prediction error is detected to conform to a continuously increasing law and there is no step impact frequency component in the corresponding vibration signal spectrum, it is determined to be a mechanical wear and decay state of the pump body, and the working condition division centroid coordinates are reversed and the neuron perception bias value of the final fully connected layer in the long short-term memory neural network unit is reduced to eliminate the physical quantity translation caused by normal mechanical aging.

[0029] Furthermore, it also includes a physical monitoring backup and reconstruction unit; the physical monitoring backup and reconstruction unit extracts dynamic time bending distance values ​​from the vibration sensor, the temperature sensor, the pressure sensor, and the current sensor to quantify the consistency of multiple parameters; when it is determined that any one of the vibration sensor, the temperature sensor, the pressure sensor, or the current sensor has a hardware channel failure, it calculates a replacement estimate based on the physical data of the other normally functioning sensors for supplementary input, and at the same time assigns zero-value weights to the feature channels corresponding to the faulty hardware in the multimodal deep learning fusion identification module to enter a degraded monitoring state. Attached Figure Description

[0030] Figure 1 A logic block diagram of an embodiment of an oil depot pump group anomaly identification system;

[0031] Figure 2 A flowchart illustrating the workflow of the adaptive threshold adjustment module provided in this embodiment of the invention;

[0032] Figure 3 This is an algorithm structure diagram of the multimodal deep learning fusion recognition module provided in an embodiment of the present invention;

[0033] Figure 4 A flowchart illustrating the workflow of the physical monitoring backup and reconstruction unit provided in this embodiment of the invention. Detailed Implementation

[0034] The following detailed description illustrates the specific implementation method:

[0035] An oil depot pump set anomaly identification system (such as...) Figure 1 As shown, it includes a multi-type data acquisition module, an adaptive threshold adjustment module, a multi-modal deep learning fusion recognition module, an abnormal trend prediction module, and an association decision support module;

[0036] The multi-type data acquisition module is used to acquire vibration signals, physical state signals, and characteristic parameters of the operating environment and load of the oil depot pump group. It converts the vibration signals into two-dimensional time-varying spectrum features and the physical state signals into standard time series data, which are then transmitted to the multimodal deep learning fusion recognition module. At the same time, the characteristic parameters are transmitted to the adaptive threshold adjustment module and the abnormal trend prediction module.

[0037] The adaptive threshold adjustment module is used to perform mean clustering calculation on historical feature parameters to obtain the centroid of the working condition division, determine the environmental working condition category according to the proximity of the currently collected feature parameters to the centroid of the working condition division, match the corresponding deviation adjustment coefficient according to the environmental working condition category, and correct the benchmark abnormal threshold based on the deviation adjustment coefficient to generate an adaptive dynamic judgment threshold and transmit it to the multimodal deep learning fusion recognition module.

[0038] The multimodal deep learning fusion recognition module is used to extract the local spatiotemporal joint features of the two-dimensional time-varying spectrogram using a convolutional neural network architecture, and to extract the dynamic temporal correlation features of the standard time series data over time using a long short-term memory network architecture. The extracted local spatiotemporal joint features and dynamic temporal correlation features are concatenated into feature vectors to obtain multimodal features. A self-attention weighting mechanism is introduced to assign sensitivity weights to each physical quantity dimension of the multimodal features for feature weighting. The weighted multimodal features are numerically compared with the adaptive dynamic judgment threshold to obtain the current anomaly probability.

[0039] The abnormal trend prediction module is used to extract the time-varying evolution trend of the standard time series data within a set sliding window scale, along with the current feature parameters, into a deep recurrent hybrid network to predict the probability curve of anomalies occurring within a set future time period.

[0040] The associated decision support module is used to input the current anomaly probability and the anomaly occurrence probability curve, use series comparison logic to retrieve and match the corresponding fault safety risk level, and interact with the task scheduling database interface to generate appropriate operation and maintenance scheduling instructions.

[0041] In practical use, at the physical level, it relies on a group of multi-dimensional physical quantity sensors deployed on key components of the oil depot pump unit. In this embodiment, the group of multi-dimensional physical quantity sensors (i.e., multi-type data acquisition modules) includes:

[0042] External piezoelectric vibration sensors (such as the Piezotronics 353Bxx series PCB vibration sensor) are typically installed at core structural points prone to mechanical resonance, such as the main bearing housing and base of the pump unit, to capture high-frequency mechanical vibration waveform signals, with the sampling frequency set to 10 kHz.

[0043] Armored temperature sensors (such as industrial platinum resistance thermometers PT100) are attached to the front / rear bearing pads or pump housing surface of the pump body to monitor temperature changes caused by conductive heat. The measurement range is -40°C to 150°C, and the measurement accuracy is ±0.2°C.

[0044] The inlet and outlet pressure probes (such as the Honeywell PX2 series) are threaded onto the pump unit's inlet and outlet oil lines to measure the pressure difference between the inlet and outlet of the oil. The range is 0-10 MPa, and the physical resolution is 0.01 MPa.

[0045] Three-phase current transformers (such as Hall effect current sensors) for motor circuits are connected in series or in a sleeve to the input three-phase lines of the drive motor power supply control cabinet to collect stator current, with a range of 0-100 A.

[0046] To achieve microsecond-level alignment of heterogeneous data, the sensors in the multi-type data acquisition modules are connected to the fieldbus controller via a bus that supports the Precision Time Protocol (PTP, IEEE 1588 standard). The PTP master clock source calculates the transmission delay through network packet exchange, thereby compensating for the time-domain sampling phase of each sensor and ensuring that the waveforms and sequences acquired by each sensor are aligned in the time dimension.

[0047] To filter out background high-frequency random white noise and electromagnetic interference, the multi-type data acquisition modules incorporate a preprocessing unit. This preprocessing unit includes a built-in bandpass digital filter with preset low-frequency and high-frequency limiting (e.g., a fourth-order Bitworth digital bandpass filter with a bandwidth designed to be 10 Hz - 5 kHz). The microampere-level current or millivolt-level voltage analog signal output from the vibration sensor is converted from analog to digital and then input into this filter to filter out power grid noise interference from the pump room and environmental non-pump body vibration noise, extracting the high-frequency vibration time-domain fluctuation component reflecting the inherent mechanical properties of the rotor.

[0048] Meanwhile, in order to eliminate the distortion mapping of the feature spatial scale distribution of the neural network by different dimensions, the preprocessing arithmetic unit inputs a mapping normalization formula for the measured values ​​of the armored temperature sensor, the inlet / outlet pressure probe, and the three-phase current transformer of the motor circuit, and maps them to the symmetrical open interval range (this embodiment takes the open interval (0,1) as an example):

[0049]

[0050] in:

[0051] : This refers to the normalized, dimensionless standard time-domain scalar data, which has no physical units;

[0052] : This represents the number of original sensor readings, which, depending on the specific sensor type, represent the measured temperature value (unit: °C), measured pressure value (unit: MPa), or measured current value (unit: A).

[0053] : This is the lower limit constant of the physical range of the corresponding sensor (e.g., the constant for PT100 is -40°C).

[0054] : This is the upper limit constant of the physical range of the corresponding sensor (for example, the constant for PT100 is 150°C).

[0055] If at a specific moment, the actual temperature value returned by the armored temperature sensor Using the above calculation logic, The lower limit and Substituting the upper limit value into the normalization formula, the corresponding value can be calculated. This standard time-domain scalar data serves as input for subsequent model scaling.

[0056] The adaptive threshold adjustment module's operating condition clustering and adjustment mechanism specifically considers the impact of operating conditions on the monitoring thresholds of each sensor. The acquisition module transmits the collected environmental and load characteristics (such as ambient temperature, humidity, and current load level) to the adaptive threshold adjustment module in real time, including ambient temperature and humidity, and load characteristics formed by the pump spindle and bearing currents.

[0057] The core logic of the adaptive threshold adjustment module (such as...) Figure 2 (As shown) This includes: First, collecting operating environment parameters for a specified historical period (e.g., the last six months) and constructing a historical feature dataset. Then, using the K-means clustering algorithm, extracting the background structure of the data in the background, defining the Euclidean geometric space constructed from the environment and load parameters, and calculating the characteristics representing different operating conditions. Centroid division for each working condition:

[0058]

[0059] in, Representing the The ambient temperature (°C) corresponding to the cluster center; Its ambient humidity (%RH); The corresponding pump set operating load level (expressed as the ratio of motor active power or motor rated maximum power).

[0060] When the system is online, the module uses the feature parameter vector collected in real time. Calculate the real-time feature vector and the centroid of each preset working condition. Cluster geometric proximity distance between them (calculated using Euclidean distance in this embodiment):

[0061]

[0062] The system selects the category corresponding to the centroid with the smallest geometric nearest distance (i.e., the largest proximity). As a current real-time environmental condition category .

[0063] For a given environmental operating condition category The module adjusts the baseline anomaly threshold set for each physical channel according to the following calculation formula to adaptively generate a dynamically changing judgment threshold in real time:

[0064]

[0065] in:

[0066] For sensors The generated real-time adaptive dynamic judgment threshold (e.g., the limiting temperature of the bearing outer ring, unit relative to the sensor) Consistent);

[0067] : Sensors that have been manually preset or calibrated under standard environmental conditions (e.g., 20°C, no load). The baseline anomaly threshold (constant);

[0068] : Assigned to the sensor Corresponding to the selected working condition category The deviation adjustment coefficient is set within the range of the two-way deviation ratio. (In this embodiment, the value is determined based on the actual situation.)

[0069] If the preset bearing outer ring temperature reference abnormal threshold is... Under heavy-load oil transportation conditions at midday in the hot summer (when the system matches the condition category to "high-temperature heavy-load category" via centroid distance), the pre-bound deviation adjustment coefficient for this category... If we get Similarly, if the matching is performed under light-load conditions on a winter night ("low-temperature light-load category"), the adjustment coefficient... The judgment threshold obtained by dynamic calculation is thus: This method successfully eliminated the masking effect of environmental thermal radiation on early-stage fault heating.

[0070] To track pump base surface temperature drift caused by seasonal changes, the adaptive threshold adjustment module rereads the newly collected historical feature set at the end of each day (e.g., 2 AM) and applies it to the clustering model. The position parameters are updated, and the deviation adjustment coefficient is recalculated after refitting. .

[0071] The feature extraction and weighted fusion mechanism of the multimodal deep learning fusion recognition module is as follows: Figure 3 As shown, the multimodal deep learning fusion recognition module includes a convolutional neural computation unit and a long short-term memory neural network unit:

[0072] The convolutional neural computation unit (CNN) performs a short-time Fourier transform (STFT, using a Hann window function with 1024 sampling points) on the filtered vibration signal over the sliding local interval, generating a two-dimensional time-varying spectrum feature reflecting the coordinated distribution of vibration energy in the time and frequency domains. The CNN is based on a CNN containing several two-dimensional convolutional kernel layers. The kernel size can be set to... This is used for sliding convolution on the image-based multi-channel features to extract the local spatial texture distribution that can characterize pitting or eccentric impact on the inner raceway of the bearing, and outputs a high-dimensional image representation vector. .

[0073] Long Short-Term Memory (LSTM) Neural Network Unit: The input is previously normalized one-dimensional sensor sequence data (including current, temperature, and pressure, etc.). It utilizes an LSTM hidden layer mechanism with self-gated input, forget, and output gates to extract dynamic temporal correlation features over a longer time span (e.g., periodic upward drift of motor operating current, cumulative pressure decrease caused by pipe resistance changes). The output is a one-dimensional time feature vector. .

[0074] Subsequently, the model concatenates the results of the two operations to obtain the extracted high-dimensional image representation vector. With one-dimensional temporal feature vector When these features are concatenated onto a unified feature axis, they are synthesized into a unified multimodal vector, which is the multimodal feature. .

[0075] These multimodal features are further input into a relevance-weighted processor. The relevance-weighted processor constructs a self-attention weight transformation network. It assigns attention weights to feature channels using the following formula:

[0076]

[0077] in:

[0078] (Query, query matrix) (Key, feature association key matrix) (Value, the true mapping matrix) are all derived from the multimodal features. Combined with learnable parameter weight matrix A low-dimensional matrix obtained through linear mapping transformation;

[0079] : Represents the channel dimension constant of the Key matrix (e.g., 64 dimensions), used for the dot product of two vectors (i.e., ... When the cross-correlation magnitude of dimensional features is known, scaling on the denominator is provided to prevent the function gradient from vanishing at high dimensions;

[0080] : is a non-linear activation function that is normalized by summation along rows, mapping the output to the probability distribution of the attention distribution formed by the row vectors (the sum of all weights is 1).

[0081] Through this transformation, the system reconstructs the allocated response gains across different sensing representation dimensions in a higher-dimensional manifold space, amplifying the sensitivity weights of vibration spectrum spikes and anomalous temperature channels. At this point, the correlation-weighted processor outputs a higher-order fusion feature incorporating spatiotemporal attention correlations. .

[0082] By using a top classifier containing a Softmax normalization layer (whose parameters are fitted during training iterations) and combining it with the previously calculated adaptive dynamic decision threshold, the probability estimate of the pump group malfunctioning at the current moment is obtained. This probability indicates the severity of the pump unit's current deviation from its healthy baseline operating trajectory.

[0083] The nonlinear modeling mechanism of the anomaly trend prediction module is as follows: the multimodal fusion model only determines the "current transient state" of the site. For "gradual degradation", the anomaly trend prediction module adopts a deep recurrent hybrid network composed of DNN (for high-dimensional static environment feature extraction) and a two-layer LSTM (for long-term spatiotemporal succession law capture).

[0084] The default width of the sliding window is set to Minutes. This prediction module continuously monitors the historical scalar time-domain sequence of physical quantities within this 60-minute sliding window sequence, while simultaneously fusing current environmental data and motor load active power. It iteratively slides forward along the time axis to calculate a specific future duration (e.g., future time) after the current time point. The probability sequence of abnormal trends at each moment within an hour is then connected to form an anomaly occurrence probability curve. This represents a dynamic mapping of the degradation trend of the pump unit.

[0085] The specific logic of the hierarchical matching and task interaction in the correlation decision support module is as follows: the correlation decision support module obtains the above-mentioned current anomaly probability. and the probability curve of abnormal occurrence Then, execute subsequent logical decisions:

[0086] This module has a built-in risk level matching rule mapping relationship. A specific level matching mapping logic is shown in Table 1 below:

[0087]

[0088] The associated decision support module connects to the pump area task scheduling information platform and the ERP material spare parts database through an open RESTful API interface. When the safety level is determined to be high-risk or medium-risk, the module will construct a structured task package containing a unique physical identifier of the pump unit, the category of the suspected source of failure (such as the bearing at the pump end), and a parameter set of estimated failure evolution time. This package is then written to the oil depot's task scheduling database using standard SQL commands or HTTP POST messages, enabling the automatic scheduling of task orders and adaptive pre-locking of spare parts.

[0089] The data interconnection platform consists of the following components and adaptive sampling frequency adjustment: the physical architecture for data interconnection and edge-cloud collaboration is as follows:

[0090] The pump station is equipped with distributed edge control nodes (e.g., rugged edge gateways with dual-core CPUs and 4GB of RAM). Sensor data is filtered and cached locally via this edge gateway. The distributed edge control nodes run the lightweight MQTT IoT communication protocol. To prevent packet loss in wireless transmissions in highly shielded areas like oil depots, the QoS (Quality of Service) transmission guarantee level is set to QoS 1 (i.e., messages are delivered at least once with an acknowledgment frame).

[0091] After the gateway aggregates the processed packet signals into timed compressed data frames and wirelessly sends them to the cloud server, the cloud server is configured with an InfluxDB distributed time-series database at the receiving layer. The database physical strategy automatically archives the data according to the master timestamp scale of the data generation and sets partitioned storage rules: the data is stored in partitioned folders divided by daily and monthly physical structures, which facilitates the establishment of a fast historical retrieval layer.

[0092] This embodiment also includes an adaptive frequency conversion acquisition controller. Its working principle and effect are as follows: the acquisition controller reads the current anomaly probability in real time. and the predicted probability curve If we are currently in a stable operating state, that is... and If the vibration signal does not exceed the safety standard, the acquisition controller determines that the local working environment is completely healthy and sends an adaptive frequency control command to the distributed edge control node. It actively reduces the physical sampling frequency of the vibration signal input (e.g., from 10 kHz to 1 kHz for intermittent sampling) and reduces the data packet frame transmission frequency of the edge node to once per minute to reduce power consumption and alleviate network burden.

[0093] If the probability is determined to be high and then stabilizes, and any value exceeds the safety standard limit, the acquisition controller immediately sends a high-speed acceleration acquisition control layer signal to the distributed edge control nodes, requiring them to adjust the sampling rate to 10kHz full-load acquisition mode. Simultaneously, the edge gateway's local FLASH memory high-speed dump mechanism is activated to prevent the loss of critical data packets due to occasional jitter in the wireless network.

[0094] The progressive offset auto-correction and normal wear calibration are specifically designed because the aging process of industrial pump sets is irreversible, causing static bias and drift in sensor signals. To distinguish between "normal aging degradation" and "sudden structural failure," the system is equipped with a progressive offset auto-correction unit, whose calculation logic is as follows:

[0095] The correction unit continuously monitors the cumulative root mean square error between the predicted values ​​generated by the online identification output of the multimodal model and the actual inspection and quality control labels. It calculates the rate of error progression of the prediction error over a set historical period.

[0096] If, after error analysis, it is determined that the error quantity exhibits a long-period, continuous, and gradual increase (e.g., the error increases linearly with a gradual slope over time), and the time-frequency graph of the vibration signal occurring at the same time is obtained, it is detected that although the total effective value increases, no low-frequency physical impact fault spectrum with physical step fracture characteristics is excited in each characteristic frequency band.

[0097] If this is the case, it can be determined that the pump body is not currently experiencing a sudden physical failure, but rather is in a normal mechanical wear-out phase due to friction.

[0098] To prevent the model's classification accuracy from being degraded, the correction unit first calculates the cumulative aging drift, and then uses this to modify the centroids of each working condition obtained by the aforementioned K-means algorithm. The geometric coordinates in space are shifted positively; simultaneously, during the background training and optimization phase, the perceptual bias values ​​of neurons in the final fully connected layer of the LSTM are reduced in reverse, and the feature weights are reconstructed. This reverse adaptive learning logic effectively eliminates static bias errors caused by normal temperature drift and normal micro-wear of the device casing.

[0099] Furthermore, considering the possibility of sensor hardware damage due to acid and alkali corrosion, high-frequency physical impacts, etc., at the oil depot site, a physical monitoring backup and reconstruction unit was also configured. The working mechanism of this unit is as follows: Figure 4 As shown.

[0100] This embodiment extracts the dynamic time warping (DTW) spatial alignment distance value between the effective value sequences of temperature, current, pressure channels and vibration time domain, and uses it to determine whether each sensor has coupling correlation in physical trend, thereby quantifying their multi-parameter consistency.

[0101] If, at a specific moment, the physical voltage signal returned by the armored temperature sensor suddenly spikes, and its characteristic value deviates drastically from the correlation range established using the DTW distance; or if the actual measured value of the armored temperature sensor returns a NaN value representing an open circuit anomaly, then the system's physical monitoring backup reconfiguration unit logic determines that the bearing armored temperature sensor hardware has suffered a damage interruption fault.

[0102] To ensure uninterrupted system operation, the reconfiguration unit then uses the currently simultaneously acquired three-phase motor current values ​​as a basis. Pump inlet and outlet pressure difference and ambient room temperature By replacing the multiple regression physical calculation formula with preset parameters, the hypothetical virtual operating temperature of the bearing is reconstructed and estimated.

[0103]

[0104] (in (These are pre-stored linear coupling regression calibration coefficients).

[0105] Replace the temperature value with the virtual reconstruction The actual physical input values ​​from the damaged sensors are then fed into the model's input level.

[0106] Simultaneously, the physical monitoring backup reconstruction unit outputs a control signal to temporarily assign zero values ​​(i.e., masking) to the weight gain of all feature convolution vectors allocated to the temperature sensor hardware channel in the self-attention mechanism network calculation of the multimodal deep learning fusion recognition module. This prevents input failure noise from affecting other normally functioning physical feature channels, allowing the system to directly switch to a degraded monitoring mode based on vibration, current, and pressure, ensuring uninterrupted operation of the anomaly recognition service.

[0107] This embodiment also discloses a method for identifying anomalies in oil depot pump sets based on the above system, including the following:

[0108] S1 Multi-Physical Source Spatiotemporal Synchronous Acquisition Steps: Vibration, temperature, pressure, and current sensors deployed at various locations in the pump group are used to align the sampling phase using the PTP time protocol and output a synchronous time-series real number stream.

[0109] S2 data in-situ preprocessing and standardization steps: edge nodes are bandpass filtered to eliminate vibration interference noise and short-time Fourier transform to convert to two-dimensional spectrum; other simulation parameters are normalized to (0,1) space using amplitude limiting mapping formula.

[0110] S3 calculation of environmental threshold adjustment envelope steps: K-means determines the environmental level and obtains the adjustment coefficient, and tunes in real time to generate an adaptive dynamic judgment threshold that includes the safety envelope curve.

[0111] S4 modal interaction fusion calculation steps: CNN processes two-dimensional vibration spectrum, LSTM processes time sequence and splices it, an attention mechanism is introduced to automatically extract sensitive feature channels, and the output represents the probability of anomaly occurrence representing the current severity level.

[0112] S5 Forward Degradation Trend Prediction Steps: Calculate the time-varying probability trend line based on the sliding window and input.

[0113] S6 safety level determination and linkage dispatch steps: The system inputs the identification results into the risk logic table. If the alarm level is exceeded, it will automatically interact with the asset task work order database to complete the maintenance and adjustment process.

[0114] The above are merely embodiments of the present invention. The invention is not limited to the fields covered by these embodiments. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are able to access all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. An anomaly identification system for oil depot pump sets, comprising a multi-type data acquisition module, an adaptive threshold adjustment module, a multi-modal deep learning fusion identification module, an anomaly trend prediction module, and an association decision support module; characterized in that: The multi-type data acquisition module is used to acquire vibration signals, physical state signals, and characteristic parameters of the operating environment and load of the oil depot pump group. It converts the vibration signals into two-dimensional time-varying spectrum features and the physical state signals into standard time series data, which are then transmitted to the multimodal deep learning fusion recognition module. At the same time, the characteristic parameters are transmitted to the adaptive threshold adjustment module and the abnormal trend prediction module. The adaptive threshold adjustment module is used to perform mean clustering calculation on historical feature parameters to obtain the centroid of the working condition division, determine the environmental working condition category according to the proximity of the currently collected feature parameters to the centroid of the working condition division, match the corresponding deviation adjustment coefficient according to the environmental working condition category, and correct the benchmark abnormal threshold based on the deviation adjustment coefficient to generate an adaptive dynamic judgment threshold and transmit it to the multimodal deep learning fusion recognition module. The multimodal deep learning fusion recognition module is used to extract the local spatiotemporal joint features of the two-dimensional time-varying spectrogram using a convolutional neural network architecture, and to extract the dynamic temporal correlation features of the standard time series data over time using a long short-term memory network architecture. The extracted local spatiotemporal joint features and dynamic temporal correlation features are concatenated into feature vectors to obtain multimodal features. A self-attention weighting mechanism is introduced to assign sensitivity weights to each physical quantity dimension of the multimodal features for feature weighting. The weighted multimodal features are numerically compared with the adaptive dynamic judgment threshold to obtain the current anomaly probability. The abnormal trend prediction module is used to extract the time-varying evolution trend of the standard time series data within a set sliding window scale, along with the current feature parameters, into a deep recurrent hybrid network to predict the probability curve of anomalies occurring within a set future time period. The associated decision support module is used to input the current anomaly probability and the anomaly occurrence probability curve, use series comparison logic to retrieve and match the corresponding fault safety risk level, and interact with the task scheduling database interface to generate appropriate operation and maintenance scheduling instructions.

2. The oil depot pump set anomaly identification system according to claim 1, characterized in that, The multi-type data acquisition module includes an external piezoelectric vibration sensor, an armored temperature sensor, an inlet pressure probe and an outlet pressure probe, and a three-phase current transformer for the motor circuit, all of which are deployed on key components of the oil depot pump group. All of the above sensors are connected to a bus that supports a high-precision time transmission protocol to align the time-domain phase of each data source.

3. The oil depot pump set anomaly identification system according to claim 2, characterized in that, The multi-type data acquisition module also includes a preprocessing arithmetic unit. The preprocessing arithmetic unit uses a preset low-frequency and high-frequency amplitude-limiting bandpass digital filter to extract the time-domain fluctuation component of the high-frequency vibration from the voltage signal output by the external piezoelectric vibration sensor. It then applies a mapping normalization formula to normalize the measured values ​​of the armored temperature sensor, the inlet pressure probe, the outlet pressure probe, and the three-phase current transformer of the motor circuit to a symmetrical open interval range.

4. The oil depot pump set anomaly identification system according to claim 3, characterized in that, The multimodal deep learning fusion recognition module includes a convolutional neural computation unit and a long short-term memory neural network unit. The convolutional neural computation unit performs local filtering convolution operations on the two-dimensional time-varying spectrum features based on short-time Fourier transform. The long short-term memory neural network unit extracts the standard time series data with a self-gated layer structure and concatenates the calculation results of the two onto a unified feature axis to complete the splicing.

5. The oil depot pump set anomaly identification system according to claim 4, characterized in that, The multimodal deep learning fusion recognition module also includes a correlation weighting processor; the correlation weighting processor converts all the spliced ​​features to the query layer, feature association key layer and real mapping value layer, performs two-vector dot product multiplication and implements amplitude scaling normalization, and reconstructs the response gain at different sensor representation components to amplify the feature intensity of the abnormal channel. The adaptive dynamic judgment threshold is obtained by multiplying the preset value of the baseline operating state by the environmental adjustment value and superimposing it. The environmental adjustment value is within the preset maximum bidirectional deviation amplitude ratio range. The adaptive threshold adjustment module performs parameter relearning of the clustering model for the environmental condition category at the end of each day and re-adjusts the deviation adjustment coefficient.

6. The oil depot pump set anomaly identification system according to claim 5, characterized in that, It also includes a data interconnection platform consisting of distributed edge control nodes located in the pump room and cloud storage resources; the distributed edge control nodes adopt a lightweight transmission protocol with transmission loss prevention verification, aggregate the processed packet signals into timed compressed data frames and send them to the cloud, and use a distributed write-only time-series database to achieve physical archiving according to the data generation time scale.

7. The oil depot pump set anomaly identification system according to claim 6, characterized in that, The data interconnection platform includes an adaptive frequency conversion acquisition controller; the adaptive frequency conversion acquisition controller compares the current anomaly probability and the anomaly occurrence probability curve with a safety standard line; if it exceeds the safety standard line, it sends an acceleration control command to the distributed edge control node to increase the sampling rate of the external piezoelectric vibration sensor and initiates a local high-speed anti-loss physical dump mechanism; if it does not exceed the safety standard line, it automatically reduces the data frame frequency to switch to a low-power intermittent scanning state.

8. The oil depot pump set anomaly identification system according to claim 7, characterized in that, It also includes a progressive offset automatic correction unit; the progressive offset automatic correction unit monitors the cumulative prediction error generated by the current abnormal probability. If the prediction error is detected to conform to the law of continuous gradual increase and there is no step impact frequency component in the corresponding vibration signal spectrum, it is determined to be the mechanical wear and decay state of the pump body, and the working condition division centroid coordinates are reversed and the sensory bias value of the neurons in the final fully connected layer of the long short-term memory neural network unit is reduced to eliminate the physical quantity translation caused by normal mechanical aging.

9. The oil depot pump set anomaly identification system according to claim 8, characterized in that, It also includes a physical monitoring backup and reconstruction unit; the physical monitoring backup and reconstruction unit extracts dynamic time bending distance values ​​from the vibration sensor, the temperature sensor, the pressure sensor, and the current sensor to quantify the consistency of multiple parameters; when it is determined that any one of the vibration sensor, the temperature sensor, the pressure sensor, or the current sensor has a hardware channel failure, it calculates a replacement estimate based on the physical data of the other normally functioning sensors for supplementary input, and at the same time assigns zero weight to the feature channel corresponding to the faulty hardware in the multimodal deep learning fusion identification module to enter a degraded monitoring state.

10. A method for identifying anomalies in oil depot pump sets, characterized in that, The system described in any one of claims 1-9 is employed.

Images