Intelligent testing method for flush valve based on multi-source sensing data

Abstract

This invention relates to the field of bathroom fixture testing technology, specifically an intelligent testing method for flush valves based on multi-source sensor data. The method includes: collecting dynamic waveforms of water pressure, flow pulse sequences, valve body vibration spectra, and sealing surface image frames of the flush valve during a test period to form a multi-source sensor data stream; performing spatiotemporal alignment and fusion processing on the multi-source sensor data stream to construct a comprehensive state field for the flush valve; and using an improved hidden Markov model constrained by the mechanical action stage characteristics of the flush valve to constrain state transition probabilities, decoding the internal working state sequence of the flush valve, and accurately locating performance anomaly events and their spatiotemporal coordinates. A set of hypotheses for anomaly causes is constructed through reverse tracing, and the root cause of the fault is determined through verification testing. This method achieves integrated analysis of multi-dimensional operating data of the flush valve, accurately identifies operational anomalies, standardizes the fault tracing process, and improves the precision of intelligent testing and fault identification of flush valves.

Description

Technical Field

[0001] This invention relates to the field of bathroom fixture testing technology, and in particular to an intelligent testing method for flush valves based on multi-source sensor data. Background Technology

[0002] Current flushing valve performance testing generally employs a single-sensor data acquisition and analysis mode, relying solely on a single physical parameter to monitor operational status. It lacks a multi-dimensional sensor information synchronous acquisition architecture, and the data acquisition scope is limited to fluid-related parameters, excluding valve body vibration signals and sealing surface image information. Various sensor data are collected and stored and analyzed independently without spatiotemporal correlation calibration and integration processing, resulting in only fragmented, localized operational data that cannot form a complete data system characterizing the overall operation of the flushing valve.

[0003] Existing testing systems cannot construct a unified comprehensive state representation of flushing valves, making it difficult to fully reproduce the actual operating conditions throughout the entire testing cycle. General hidden Markov models are directly applied to decode the internal working states of flushing valves without considering the inherent attributes of the valve's mechanical action stages to set adaptation rules. State transition probabilities lack specific constraints, resulting in deviations between the analyzed working state sequences and actual operating conditions. Traditional detection methods cannot accurately identify performance anomalies, pinpoint the spatiotemporal location of anomalies, and lack standardized data tracing logic for fault diagnosis, making it difficult to locate the root cause of faults through systematic deduction. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent testing method for flushing valves based on multi-source sensor data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent testing method for flushing valves based on multi-source sensor data, comprising: Acquire multi-source sensor data streams of the flushing valve during the test cycle. The multi-source sensor data streams include dynamic waveforms of water pressure, flow pulse sequences, valve body vibration spectrum, and sealing surface image frames. The multi-source sensor data stream is spatiotemporally aligned and fused to generate a comprehensive state field of the flushing valve within the test cycle. Based on the comprehensive state field of the flushing valve, an improved hidden Markov model is used to decode the internal working state sequence of the flushing valve. The improved hidden Markov model constrains the state transition probability based on the mechanical action stage characteristics of the flushing valve. Based on the decoded internal working state sequence, locate the abnormal performance events of the flushing valve and the corresponding spatiotemporal coordinates of the events; Based on the performance anomaly events and their corresponding spatiotemporal coordinates, a set of hypotheses for the causes of the anomalies is generated by reverse tracing, and the set of hypotheses for the causes of the anomalies is verified and tested to determine the root cause of the fault.

[0006] As a further aspect of the present invention, the multi-source sensor data stream is subjected to spatiotemporal alignment and fusion processing to generate a comprehensive state field of the flushing valve within the test period, including: The dynamic waveform of water pressure is subjected to a segmented stability test, which divides the pressure stabilization stage, the pressure rise stage, and the pressure fall stage. The peak value and pulse width are simultaneously measured on the flow pulse sequence to identify the start time, peak time, end time and pulse intensity of each flow pulse; Time-frequency analysis was performed on the vibration spectrum of the valve body to extract the energy change curves of the characteristic frequency bands corresponding to valve core impact, spring reset, and sealing surface friction over time. Perform time-series analysis on the image frames of the sealing surface, calculate the pixel gradient changes in the sealing area in each frame, and generate a sealing surface deformation sequence; A unified test time axis is established, which maps the divided pressure stabilization phase, pressure rise phase, and pressure fall phase, the start time, peak time, end time, and pulse intensity of the identified flow pulse, the energy change curve of the extracted characteristic frequency band over time, and the generated sealing surface deformation sequence to the unified test time axis. On the unified test time axis, for each time slice, the pressure stage label, flow pulse parameter, vibration energy value and deformation value mapped to this slice are tensor spliced ​​to form a comprehensive state vector of the time slice. The combined state vectors of all time slices are combined in chronological order to form the combined state field of the flushing valve.

[0007] As a further aspect of the present invention, the step of decoding the internal operating state sequence of the flushing valve using an improved hidden Markov model includes: The improved hidden Markov model contains a set of hidden states, which correspond one-to-one with the physical working stages of the flushing valve, including the water inlet pressurization state, valve open state, high-speed flushing state, valve closed state, and reset sealing state. The improved Hidden Markov Model constrains the state transition probability based on the mechanical action stage characteristics of the flushing valve. Specifically, the state transition is only allowed to proceed unidirectionally in the order of water inlet pressure storage state, valve open state, high-speed flushing state, valve closed state, and reset sealing state, or remain in the current state. The integrated state vector of each time slice in the integrated state field of the flushing valve is used as the observation of the improved hidden Markov model at the current time step. Using the Viterbi algorithm, we find the hidden state sequence under a given observation sequence, which is the decoded internal working state sequence; During the decoding process, the observation probability distribution of the water inlet pressure storage state and the valve closed state is weighted and adjusted based on the stage division results of the water pressure dynamic waveform.

[0008] As a further aspect of the present invention, the Viterbi algorithm is used to find the hidden state sequence under a given observation sequence, wherein the hidden state sequence is the decoded internal working state sequence, including: The comprehensive state field of the flushing valve is arranged in the order of the test time axis to form an observation sequence, and the state transition probability matrix, observation probability matrix and initial state probability vector of the improved hidden Markov model are used as known parameters. During the initialization phase of the Viterbi algorithm, the initial probability of each hidden state corresponding to the current observation value at the first time slice is calculated, and the probability value of each state and the previous state being empty are recorded. In the recursive phase of the Viterbi algorithm, the observation sequence is traversed backward from the second time slice. For each hidden state in the current time slice, the probability of transitioning from each state in the previous time slice to the current state and generating the current observation value is calculated. The path with the highest probability is selected from all transition paths, and the maximum probability value and the corresponding previous state are recorded. In the termination phase of the Viterbi algorithm, after traversing all observations, the state with the highest probability is selected from all states of the last time slice as the endpoint of the optimal path. Starting from the end of the optimal path, based on the previous state information recorded in the recursive stage, backtrack to the first time slice, and the resulting hidden state sequence is the hidden state sequence under the given observation sequence. The hidden state sequence is output as the decoded internal working state sequence.

[0009] As a further aspect of the present invention, the step of locating the performance abnormality event of the flushing valve and the corresponding spatiotemporal coordinates of the event based on the decoded internal working state sequence includes: The decoded internal working state sequence is compared with the standard working state sequence, which defines the standard duration range and standard transition relationship of each hidden state under normal working conditions. Identify state timeout events or state insufficiency events in the internal working state sequence whose state duration exceeds the standard duration range; Identify state transition events or state missing events in the internal working state sequence that violate the standard transition relationship; In the overall state field of the flushing valve, find the time slices corresponding to the state timeout event, state insufficiency event, state jump event, or state missing event. From the comprehensive state vector corresponding to the time slice found, extract the abnormal pressure stage label, flow pulse parameter, vibration energy value and deformation value, and encapsulate the extracted abnormal parameters and event type to form the performance abnormal event. The position of the performance anomaly event and its corresponding time slice on the test time axis are collectively defined as the spatiotemporal coordinates of the event.

[0010] As a further aspect of the present invention, based on the performance anomaly event and its corresponding spatiotemporal coordinates, a set of hypotheses for the cause of the anomaly is generated through reverse tracing, including: Establish a knowledge graph for flush valve failure, wherein the knowledge graph uses performance abnormality events as nodes, component failure causes as nodes, and causal links between events and causes as edges; The performance abnormality events are input into the flushing valve fault knowledge graph for matching and retrieval. Based on the matched causal links, the source of one or more candidate component failures leading to the performance anomaly event is traced, forming an initial set of anomaly cause hypotheses; Based on the spatiotemporal coordinates corresponding to the event, a multi-source sensor data subfield within the time window preceding the occurrence of the abnormal event is extracted from the comprehensive state field of the flushing valve. Analyze the changing trends and collaborative relationships of different sensor data in the multi-source sensor data subfield, and score the support of each candidate component failure cause in the initial set of anomaly cause hypotheses. The candidate component failure causes are sorted and filtered based on the support scores to generate a set of anomaly cause hypotheses.

[0011] As a further aspect of the present invention, the set of hypotheses for the causes of the anomalies is verified and tested to determine the root cause of the fault, including: For each candidate component failure cause in the set of anomaly cause hypotheses, a targeted verification test plan is designed. The targeted verification test plan includes applying a specific excitation signal, acquiring higher-precision local sensing data, or executing a specific action sequence. On the flushing valve test platform, each of the aforementioned targeted verification test schemes was executed in sequence, and high-precision response data was collected synchronously during the verification test. The collected high-precision response data is compared with the expected response data of the corresponding candidate component failure cause under the standard failure mode; Calculate the consistency metric for each comparison, which integrates waveform similarity, feature matching degree, and time sequence alignment degree; Candidate component failure causes whose consistency metric values ​​exceed a preset confidence threshold are identified as verified root causes of failure. If the consistency metric values ​​of all candidate component failure causes do not exceed the preset confidence threshold, then the performance anomaly event and the corresponding multi-source sensor data subfield are marked as unknown anomaly mode.

[0012] As a further aspect of the present invention, the high-precision response data collected is compared with the expected response data of the corresponding candidate component failure cause under the standard failure mode, including: For candidate component failure causes corresponding to valve core wear failure, extract the valve core motion acceleration waveform and sealing surface leakage acoustic emission signal from the high-precision response data; The extracted valve core motion acceleration waveform and the expected acceleration waveform under the standard valve core wear fault mode are dynamically time-warped, and the path distance between the valve core motion acceleration waveform and the standard valve core wear fault mode is calculated as the waveform difference degree. The extracted acoustic emission signal from the sealing surface leakage was subjected to spectral analysis, and the ratio between its dominant frequency component energy and the dominant frequency component energy of the standard leakage spectrum was calculated as the spectral fit. By combining the waveform difference and the spectrum matching, a consistency metric for valve core wear faults is obtained through weighted summation.

[0013] As a further aspect of the present invention, the method further includes: After determining the root cause of all performance anomalies in the current test cycle, a test analysis report is generated. The test analysis report records the spatiotemporal coordinates of each performance anomaly, the set of anomaly cause hypotheses, the verification test process, and the final result of the root cause determination. The test analysis report is associated and stored with the corresponding multi-source sensor data stream, flushing valve integrated state field and internal working state sequence to form a complete test case; The complete test cases are added to the historical case knowledge base to optimize the parameters of the improved hidden Markov model, enrich the causal links of the flushing valve fault knowledge graph, and update the expected response data under standard fault modes.

[0014] As a further aspect of the present invention, the complete test cases are added to the historical case knowledge base to optimize the parameters of the improved hidden Markov model, including: From the historical case knowledge base, select all complete test cases marked with a definite root cause of the failure; Extract the overall state field and internal working state sequence of the flushing valve from each selected complete test case; Use all extracted flush valve integrated state fields as the new observation sequence training set, and use all extracted internal working state sequences as the new state sequence label set; A supervised learning algorithm is used to re-estimate the observation probability matrix of the improved Hidden Markov Model using the new observation sequence training set and the new state sequence label set. During the re-estimation process, abnormal observations that occur frequently in the root cause cases of failure are given higher probability weights in the observation probability matrix, thereby optimizing the model's ability to identify abnormal states.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This system integrates multiple types of sensor data, including dynamic waveforms of water pressure, flow pulse sequences, valve body vibration spectra, and sealing surface image frames, to achieve spatiotemporal alignment and fusion of multi-source sensor data streams. It unifies the temporal reference and spatial representation dimensions of different sensor acquisition terminals. This resolves the data fragmentation issues caused by inconsistent sampling frequencies and misaligned timing nodes among various sensor data, integrating heterogeneous discrete sensor information into a unified state representation. It comprehensively covers the fluid operation, mechanical vibration, and sealing structure characteristics of the flushing valve, achieving a complete and holistic representation of the operating state within the test cycle, thus overcoming the inherent limitation of a single parameter only reflecting local operating characteristics.

[0016] By combining the inherent operational patterns of the flushing valve's mechanical action stages, specific constraints are imposed on the state transition probabilities of the Hidden Markov Model (HMM) to adapt to the stage-specific changes in the actual mechanical operation of the flushing valve, thus optimizing the internal logic of the model's state deduction. Based on the optimized model, the internal working state sequence is accurately decoded, and performance anomaly-related change information is captured using the complete state sequence, matching the temporal and spatial location information corresponding to the abnormal events. Reverse tracing is conducted based on the anomaly's spatiotemporal information to construct a complete set of hypotheses for the anomaly's causes. The root cause of the fault is identified through a standardized verification process. The entire process relies on data-driven logic to advance anomaly identification and fault diagnosis, moving away from judgment methods dominated by human experience and improving the automated testing and fault identification process system for flushing valves. Attached Figure Description

[0017] Figure 1 This is a flowchart of the intelligent testing method for flushing valves based on multi-source sensor data described in this invention; Figure 2 A flowchart for decoding the internal working state sequence of an improved Hidden Markov Model; Figure 3 A flowchart for locating performance anomaly events and their spatiotemporal coordinates. Detailed Implementation

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

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1 This invention provides an intelligent testing method for flushing valves based on multi-source sensor data. The overall implementation scheme is as follows: The system acquires multi-source sensor data streams generated by the flushing valve throughout the entire test cycle. These streams integrate various sensor information, including dynamic water pressure waveforms, flow pulse sequences, valve body vibration spectra, and sealing surface image frames. Subsequently, these heterogeneous data streams from different sensors with varying sampling rates are aligned and fused temporally and spatially to construct a comprehensive state field that fully and synchronously reflects the state of each component of the flushing valve at every moment along the test timeline. Based on this comprehensive state field, an improved Hidden Markov Model (HMM) is used to decode the internal, unobservable working state sequence of the flushing valve. The improvement lies in the strict constraint of its state transition probability matrix on the mechanical action stages that the flushing valve must follow (such as the sequence of water inlet, opening, flushing, closing, and reset), making the decoding more consistent with physical laws. The decoded internal working state sequence reveals the actual working process of the flushing valve. By comparing this sequence with a preset standard working state sequence, performance anomalies such as abnormal state duration or abnormal state transitions can be accurately located, and the occurrence times of these events along the test timeline can be determined. Furthermore, for each identified anomaly, by combining detailed features from multi-source sensor data before and after its occurrence, reverse causal reasoning is performed from a pre-constructed fault knowledge graph to generate a set of candidate component failure causes that may lead to the anomaly—a set of anomaly cause hypotheses. Finally, targeted verification tests are designed and executed for each candidate cause in this hypothesis set. By applying specific stimuli, collecting high-precision local response data, and comparing it with the expected response under standard failure modes, the root cause of the initial performance anomaly is verified and ultimately determined.

[0021] In one embodiment of the present invention, the dynamic waveform of water pressure is subjected to segmented stationarity testing, which aims to divide it into three distinct pressure stabilization, pressure rise, and pressure fall stages. Simultaneously, peak value detection and pulse width measurement are performed on the flow pulse sequence to identify the start time, peak time, end time, and pulse intensity of each flow pulse. Time-frequency analysis is performed on the valve body vibration spectrum to extract the energy-time curves of characteristic frequency bands corresponding to key mechanical events such as valve core impact, spring reset, and sealing surface friction. Time-series analysis is performed on the sealing surface image frames, and by calculating the pixel gradient changes in the sealing area of ​​each frame, a sealing surface deformation sequence reflecting the morphological changes of the sealing contact surface is generated. A unified test time axis is established, mapping the labels of the aforementioned pressure stabilization, pressure rise, and pressure fall stages, the identified start time, peak time, end time, and pulse intensity parameters of the flow pulses, the extracted characteristic frequency band energy curves, and the generated sealing surface deformation sequence onto this unified time axis. On this unified test timeline, for each time slice with equal or unequal intervals, the pressure stage label, flow pulse parameters, vibration energy value, and deformation value mapped to that slice are tensor-stitched together to form a comprehensive state vector representing the global state of the flushing valve at that moment. Combining all these comprehensive state vectors of time slices in chronological order constitutes a comprehensive state field of the flushing valve that spans the entire test cycle and integrates multi-dimensional information.

[0022] In practical implementation, the multi-source sensor data streams undergo spatiotemporal alignment and fusion processing to generate a comprehensive state field for the flushing valve. During the flushing valve test cycle, pressure sensors deployed in the water supply pipeline acquire dynamic water pressure waveforms, flow meters deployed in the pipeline acquire flow pulse sequences, accelerometers installed on the valve body housing acquire vibration signals and generate valve body vibration spectra through Fourier transform, and high-speed industrial cameras facing the sealing pair acquire image frames of the sealing surface. These data collectively constitute the multi-source sensor data stream. In practical implementation, the dynamic water pressure waveform undergoes segmented stationarity testing. By calculating the rate of change of the root mean square value of the signal within the sliding time window and setting a threshold, the waveform is identified into stationary segments where the absolute value of the slope is lower than the first threshold, rising segments where the slope is continuously positive, and falling segments where the slope is continuously negative, thereby dividing the pressure into a stable stage, a pressure rising stage, and a pressure falling stage. In an example of a flushing cycle, the pressure stabilization phase corresponds to the period when the water supply pressure is maintained at 400 kPa, the pressure rise phase corresponds to the period when the pressure rises from 400 kPa to 600 kPa during the valve closing process, and the pressure fall phase corresponds to the period when the pressure drops rapidly from 600 kPa to 100 kPa after the valve is opened.

[0023] In some embodiments, peak detection and pulse width measurement are performed simultaneously on the flow pulse sequence to identify the start time, peak time, end time, and pulse intensity of each flow pulse. The differential threshold method is used to determine the start time as the moment the flow signal exceeds the baseline noise at its first rising edge, the peak time as the moment the flow signal reaches its global maximum value, and the end time as the moment the flow signal falls back below the baseline noise. The pulse intensity is the maximum flow rate within that pulse period. In a specific example, a complete flushing flow pulse starts at 2.1 seconds, peaks at 3.0 seconds, ends at 4.0 seconds, and has a pulse intensity of 15.0 L / min. Optionally, time-frequency analysis is performed on the valve body vibration spectrum. Short-time Fourier transform is used to obtain the distribution of vibration energy on the time-frequency plane, extracting the energy variation curves over time for the 80-120Hz frequency band corresponding to valve core impact, the 200-300Hz frequency band corresponding to spring reset, and the 500-800Hz frequency band corresponding to friction with the sealing surface. In practical implementation, temporal analysis is performed on the image frames of the sealing surface. The gradient magnitude of each pixel within a predefined annular sealing region in each frame is calculated. The arithmetic mean of the gradient magnitudes of all pixels is then calculated to generate a scalar sequence characterizing the overall microscopic deformation degree of the sealing surface, i.e., the sealing surface deformation sequence. The calculation formula for the sealing surface deformation sequence can be expressed as: in: This represents the deformation value of the sealing surface at time t. It is the total number of pixels in the sealed area. At time t, the image frame is at pixel coordinates grayscale value at that location It is the gradient operator. This indicates the calculation of the magnitude of the gradient vector.

[0024] In practical implementation, a unified test time axis is established, with the test start point as the zero point. Its time resolution is consistent with the sampling interval of the sensor with the highest sampling frequency in the multi-source sensor data stream. The time interval labels for the defined pressure stabilization, pressure rise, and pressure fall phases, the start time, peak time, end time, and pulse intensity parameters of each identified flow pulse, the discrete data points on the energy-time curve of the extracted characteristic frequency bands, and each deformation value and its timestamp in the generated sealing surface deformation sequence are all mapped onto the unified test time axis using linear interpolation or nearest neighbor matching. On the unified test time axis, for each equally spaced time slice, the pressure phase label, flow pulse parameters, vibration energy value, and deformation value mapped to that time slice are tensor-concatenated to form a comprehensive state vector for the time slice. The comprehensive state vector for the time slice is a multi-dimensional vector, its dimension equal to the encoding dimension of the pressure phase label, the number of flow pulse parameters, the number of vibration energy values, and the number of deformation values. The combined state vectors of all time slices are combined in chronological order to form the flushing valve's comprehensive state field. This comprehensive state field is a two-dimensional matrix, where rows correspond to time slices and columns correspond to the various feature dimensions of the comprehensive state vectors. In essence, through these steps, the originally heterogeneous and asynchronous multi-source sensor data are fused into a flushing valve's comprehensive state field that is strictly aligned in time and uniformly represented in features, providing a structured input for subsequent state decoding.

[0025] In one embodiment of the present invention, see [reference] Figure 2The improved Hidden Markov Model (HMM) defines a set of hidden states, each corresponding to a physical operating stage of the flushing valve. Typical states include inlet pressurization, valve opening, high-speed flushing, valve closing, and reset sealing. The model constrains state transition probabilities based on the mechanical characteristics of the flushing valve's operating stages. Specifically, state transitions are only allowed in a unidirectional sequence: inlet pressurization, valve opening, high-speed flushing, valve closing, and reset sealing. Alternatively, the model can remain in its current state, but reverse or skip-transition state transitions are prohibited. The composite state vector of each time slice in the flushing valve's composite state field is considered the observation value of the improved HMM at the current time step. The Viterbi algorithm is used to find the most probable sequence of hidden states given the observation sequence; this sequence is the decoded internal operating state sequence. The execution of the Viterbi algorithm includes initialization, recursion, and termination phases. In the initialization phase, the initial probability corresponding to each hidden state and the current observation value at the first time slice is calculated, and the probability value of each state and the previous state (empty) are recorded. In the recursive phase, starting from the second time slice, the observation sequence is traversed backward. For each hidden state in the current time slice, the probability of transitioning from each state in the previous time slice to the current state and generating the current observation value is calculated. The path with the highest probability is selected from all possible transition paths, and this maximum probability value and the corresponding previous state are recorded. In the termination phase, after traversing all observation values, the state with the highest probability is selected from all states in the last time slice as the endpoint of the optimal path. Starting from the endpoint of the optimal path, based on the previous state information recorded in the recursive phase, the model backtracks to the first time slice, and the resulting hidden state sequence is the decoding result. During the decoding process, the observation probability distribution of the inlet pressure storage state and the valve closing state is weighted and adjusted in conjunction with the stage division results of the water pressure dynamic waveform. For example, when the pressure label corresponding to the observation value is "pressure stable stage", the probability of the model observing the inlet pressure storage state is increased.

[0026] In practical implementation, an improved Hidden Markov Model (HMM) is used to decode the internal working state sequence of the flushing valve. The improved HMM contains a set of hidden states, each corresponding one-to-one with a physical working stage of the flushing valve. These hidden states include the inlet pressurization state, valve opening state, high-speed flushing state, valve closing state, and reset sealing state. The improved HMM constrains the state transition probabilities based on the mechanical action stage characteristics of the flushing valve. Specifically, state transitions are only allowed to proceed unidirectionally in the order of inlet pressurization state, valve opening state, high-speed flushing state, valve closing state, and reset sealing state, or remain in the current state. This constraint is achieved by defining a state transition probability matrix, where only sequential transitions or self-transitions are allowed to have non-zero probabilities, while reverse or jump transition probabilities are forced to zero. The comprehensive state vector of each time slice in the comprehensive state field of the flushing valve serves as the observation value of the improved HMM at the current time step.

[0027] In some embodiments, the Viterbi algorithm is used to find the most probable sequence of hidden states under a given observation sequence. This sequence is the decoded internal working state sequence. The input to the Viterbi algorithm is the sequence of observations, i.e., the integrated state field of the flushing valve arranged in the order of the test time axis, and the known parameters of the improved Hidden Markov Model, including the state transition probability matrix, the observation probability matrix, and the initial state probability vector. In the initialization phase of the Viterbi algorithm, the initial probability corresponding to each hidden state and the current observation value at the first time slice is calculated, and the probability value of each state and the previous state being empty are recorded. For the first time slice, the probability value of each hidden state is calculated by multiplying the initial state probability vector by the probability component of the corresponding observation value in the observation probability matrix. In the recursive phase of the Viterbi algorithm, the observation sequence is traversed backward from the second time slice. For each hidden state in the current time slice, the probability of transitioning from each state in the previous time slice to the current state and generating the current observation value is calculated. The path with the highest probability is selected from all transition paths, and the maximum probability value and the corresponding previous state are recorded. The calculation of the recursive process can be expressed as: in: This indicates that the hidden state at time t is The maximum probability among all paths, It is the total number of hidden states. From state Transition to state The probability, It is in state The following observations The probability, This is the comprehensive state vector at time t. In the termination phase of the Viterbi algorithm, after traversing all observations, the state with the highest probability from all states in the last time slice is selected as the endpoint of the optimal path. Starting from the endpoint of the optimal path, based on the previous state information recorded in the recursive phase, we backtrack to the first time slice. The resulting hidden state sequence is the hidden state sequence under the given observation sequence. The hidden state sequence is output as the decoded internal working state sequence.

[0028] In specific implementation, the decoding process incorporates the stage division results of the water pressure dynamic waveform, and weights the observation probability distributions of the inlet pressure storage state and the valve closing state. Optionally, when the pressure stage label mapped to the time slice is "pressure stable stage," the observation probability component corresponding to the inlet pressure storage state is multiplied by a weight coefficient greater than 1, such as 1.5, to enhance the model's tendency to select the inlet pressure storage state under the current observation. Similarly, when the pressure stage label mapped to the time slice is "pressure rising stage," the observation probability component corresponding to the valve closing state is multiplied by a weight coefficient greater than 1, such as 1.5. In some embodiments, the observation probability matrix is ​​obtained through supervised training using a large amount of flushing valve comprehensive state field data from normal flushing cycles and manually labeled internal working state sequences. Its dimension is the number of hidden states multiplied by the number of discretized intervals of the observation feature vector. The improved Hidden Markov Model, through the Viterbi algorithm-decoded internal working state sequence, can map continuous, multi-dimensional fused observation data into a discrete state sequence conforming to mechanical temporal logic. It is understandable that, for example, a normal decoding result sequence may present as "water inlet pressure storage state, water inlet pressure storage state, valve open state, high-speed flushing state, high-speed flushing state, valve closed state, valve closed state, reset sealing state", while an abnormal decoding result may present a timing error of "water inlet pressure storage state, valve open state, water inlet pressure storage state".

[0029] In one embodiment of the present invention, see [reference] Figure 3The decoded internal operating state sequence is compared with a standard operating state sequence. The standard operating state sequence defines the standard duration range of each hidden state under normal operating conditions and the allowed standard transition relationships between states. Through comparison, state timeout events or state insufficiency events in the internal operating state sequence that exceed the standard duration range are identified. Simultaneously, state jump events or state missing events that violate the standard transition relationships in the internal operating state sequence are identified. In the flushing valve's comprehensive state field, time slices corresponding to the identified state timeout events, state insufficiency events, state jump events, or state missing events are located. From the comprehensive state vectors corresponding to these time slices, abnormal pressure stage labels, flow pulse parameters, vibration energy values, and deformation values ​​are extracted. These extracted abnormal parameters are encapsulated together with the event type to form a structured description of a performance anomaly event. The position of this performance anomaly event and its corresponding time slice on the test time axis is defined as the spatiotemporal coordinates of the event. A knowledge graph for flushing valve failure is established, with various performance anomaly events as nodes, possible component failure causes as nodes, and causal reasoning links between events and causes as edges. Encapsulated performance anomaly events are input into the knowledge graph for matching and retrieval. Based on the matched causal links, one or more candidate component failure causes leading to the performance anomaly event are traced backward, forming an initial set of anomaly cause hypotheses. Combining the spatiotemporal coordinates of the event, a multi-source sensor data subfield within the time window preceding the anomaly event is extracted from the flushing valve's overall state field. The changing trends and synergistic relationships of different sensor data within this multi-source sensor data subfield are analyzed, and a support score is assigned to each candidate component failure cause in the initial set of anomaly cause hypotheses. Based on the support scores, the candidate component failure causes are sorted and filtered to generate the final set of anomaly cause hypotheses.

[0030] In practical implementation, the intelligent testing method for flushing valves based on multi-source sensor data involves locating performance anomaly events and their corresponding spatiotemporal coordinates based on the decoded internal operating state sequence, and then using this as a basis to reverse-engineer the cause hypothesis set. The decoded internal operating state sequence is compared with a standard operating state sequence, which defines the standard duration range of each hidden state under normal operating conditions and the permissible standard transition relationships between states. The standard operating state sequence is predefined in the form of a state transition diagram or state-timetable, and the standard duration range is usually obtained based on statistics from a large number of normal flushing cycles. A specific definition of a standard operating state sequence is shown in Table 1.

[0031] Table 1 Standard Operating State Sequence Definition Table: In some embodiments, comparisons are used to identify state timeout events or state insufficiency events in the internal working state sequence whose state duration exceeds the standard duration range. For example, in a test, if the decoded internal working state sequence shows that the "valve open state" lasts for 0.8 seconds, exceeding the upper limit of the standard duration range by 0.3 seconds, a state timeout event is identified. Comparisons are also used to identify state jump events or state missing events in the internal working state sequence whose state transition relationships violate the standard transition relationships. For example, if the decoded sequence shows a direct jump from "water inlet pressurization state" to "high-speed flushing state," skipping the "valve open state," a state jump event is identified; if the sequence lacks the "reset sealing state," a state missing event is identified.

[0032] In practical implementation, within the comprehensive state field of the flushing valve, time slices corresponding to state timeout events, state insufficiency events, state transition events, or state missing events are located. For state timeout events, all time slices corresponding to the duration of the abnormal state are located; for state transition events, the time slice corresponding to the transition moment in which the state transition occurs is located. From the comprehensive state vector corresponding to the found time slices, the abnormal pressure stage label, flow pulse parameters, vibration energy value, and deformation value are extracted. The extracted abnormal parameters and event type are encapsulated together to form a performance abnormality event. The encapsulated data structure of the performance abnormality event includes an event type field, an occurrence time field, and a feature vector containing specific abnormal sensor readings. The position of the performance abnormality event and its corresponding time slice on the test time axis is defined as the spatiotemporal coordinates corresponding to the event. The spatiotemporal coordinates include the test cycle identifier, the start timestamp of the abnormality, and the end timestamp.

[0033] In some embodiments, a set of hypotheses about the causes of anomalies is generated by reverse tracing based on performance anomaly events and their corresponding spatiotemporal coordinates. A flushing valve fault knowledge graph is established, with performance anomaly events as nodes, component failure causes as nodes, and causal links between events and causes as edges. The edges of the knowledge graph are accompanied by causal strength weights and prior probabilities. The encapsulated performance anomaly events are input into the flushing valve fault knowledge graph for matching and retrieval. Based on the matched causal links, one or more candidate component failure causes leading to the performance anomaly events are traced, forming an initial set of hypotheses about the causes of anomalies. For example, for the event "valve opening state timeout," the knowledge graph may trace back to multiple candidate component failure cause nodes such as "valve core movement jamming," "drive spring fatigue," and "waterway blockage."

[0034] Optionally, combining the spatiotemporal coordinates corresponding to the event, a multi-source sensor data subfield within a time window preceding the occurrence of the abnormal event is extracted from the comprehensive state field of the flushing valve. The length of the time window is configurable. The changing trends and collaborative relationships of different sensor data within the multi-source sensor data subfield are analyzed, and a support score is assigned to each candidate component failure cause in the initial set of anomaly cause hypotheses. The support score function comprehensively considers the degree of matching between the characteristics of each sensor signal and the typical patterns of failure causes. The formula for calculating the support score can be expressed as: , in: This indicates the support level for the cause of failure of the candidate component. It is the number of feature types in multi-source sensor data. It is the weight of the first type of feature. It is a similarity calculation function. It is the type-i eigenvalue observed from the multi-source sensor data subfield. This represents the expected type-1 feature value under typical failure cause patterns. It can be understood that candidate component failure causes are sorted and filtered based on support scores, for example, retaining only candidate causes with scores above a threshold, to generate the final set of anomaly cause hypotheses. The anomaly cause hypothesis set is an ordered list, where each item contains a description of a candidate component failure cause and its calculated support score.

[0035] In one embodiment of the present invention, a targeted verification test plan is designed for each candidate component failure cause in the set of abnormal cause hypotheses. The targeted verification test plan includes applying a specific excitation signal, acquiring higher-precision local sensor data, or executing a specific action sequence. On the flushing valve test platform, each targeted verification test plan is executed sequentially, and high-precision response data is acquired synchronously during the verification test. The acquired high-precision response data is compared with the expected response data of the corresponding candidate component failure cause under the standard failure mode. A consistency metric is calculated for each comparison, which integrates multiple dimensions such as waveform similarity, feature matching degree, and time sequence alignment. For the candidate component failure cause corresponding to valve core wear failure, the comparison process specifically includes: extracting the valve core motion acceleration waveform and the sealing surface leakage acoustic emission signal from the high-precision response data; dynamically time-warping the extracted valve core motion acceleration waveform with the expected acceleration waveform under the standard valve core wear failure mode, and calculating the warped path distance between the two as the waveform difference; and performing spectral analysis on the extracted sealing surface leakage acoustic emission signal, calculating the ratio between its dominant frequency component energy and the dominant frequency component energy of the standard leakage spectrum as the spectral fit. By combining waveform difference and spectral consistency, a consistency metric for valve core wear faults is calculated through weighted summation. Candidate component fault causes whose consistency metric values ​​exceed a preset confidence threshold are identified as verified root causes. If the consistency metric values ​​of all candidate component fault causes do not exceed the preset confidence threshold, the original performance anomaly event and its corresponding multi-source sensor data subfield are marked as an unknown anomaly mode.

[0036] In practical implementation, the set of hypotheses for abnormal causes is validated through testing to determine the root cause of the failure. For each candidate component failure cause in the set of hypotheses, a targeted validation test plan is designed, the specific content of which is customized based on the physical mechanism of the candidate component failure cause. For example, for the candidate cause "valve core wear," the targeted validation test plan includes applying an excitation vibration signal containing a specific frequency component to the valve body and simultaneously acquiring the valve core motion acceleration waveform and the acoustic emission signal of the sealing surface leakage; for the candidate cause "sealing ring aging," the targeted validation test plan includes performing multiple consecutive low-pressure flushing actions and acquiring the sealing surface contact force distribution data during each flushing process; for the candidate cause "drive spring fatigue," the targeted validation test plan includes applying a reverse water pressure pulse during the valve closing phase and acquiring the spring return speed versus displacement curve. The targeted validation test plan aims to excite or amplify the characteristic response of a specific failure mode. On the flushing valve test platform, each targeted validation test plan is executed sequentially, and high-precision response data is acquired simultaneously during the validation test. The sampling rate of the high-precision response data is higher than the sampling rate of the initial multi-source sensor data stream.

[0037] In some embodiments, the collected high-precision response data is compared with the expected response data of the corresponding candidate component failure cause under standard failure modes. The expected response data under standard failure modes comes from a failure mode library, which stores high-precision response data templates generated by various typical failures under standard test conditions. A consistency metric is calculated for each comparison, which combines waveform similarity, feature matching, and time alignment. The consistency metric can be expressed as the following formula: , in: Represents a consistency metric. This represents the normalized dynamic time-warped path distance, used to measure waveform similarity. The spectral correlation coefficient is used to measure the degree of feature matching. Indicates the time alignment of key event points. These are the weighting coefficients for each term, and they satisfy... .

[0038] In practical implementation, for candidate component failure causes corresponding to valve core wear failure, the valve core motion acceleration waveform and sealing surface leakage acoustic emission signal are extracted from high-precision response data. The extracted valve core motion acceleration waveform is dynamically time-normalized with the expected acceleration waveform under the standard valve core wear failure mode. The normalized path distance between the two waveforms is calculated, and this path distance is divided by the reference length to obtain the waveform difference. The extracted sealing surface leakage acoustic emission signal is subjected to spectral analysis, and the ratio between its dominant frequency component energy and the dominant frequency component energy of the standard leakage spectrum is calculated as the spectral consistency. Combining the waveform difference and spectral consistency, a consistency metric for valve core wear failure is calculated through weighted summation. See Table 2, which shows a verification test comparison result: Table 2 Comparison Table of Valve Core Wear Failure Verification Test: It is understood that candidate component failure causes whose consistency metric values ​​exceed a preset confidence threshold are identified as the verified root causes of failure. The preset confidence threshold is determined statistically from historical verification data, for example, set to 0.85. If the consistency metric values ​​of all candidate component failure causes do not exceed the preset confidence threshold, the initial performance anomaly event and its corresponding multi-source sensor data subfield are marked as an unknown anomaly mode, and the data for the unknown anomaly mode will be stored for subsequent analysis. In some embodiments, when the consistency metric values ​​of multiple candidate causes all exceed the threshold, the candidate component failure cause with the highest consistency metric value is selected as the primary root cause of failure. Optionally, the verification test process and the determination results are recorded and a report is generated. It is understood that by executing targeted verification test plans and performing high-precision data comparison, the fundamental failure cause leading to the initial performance anomaly can be screened and confirmed from the set of anomaly cause hypotheses, thereby completing a closed-loop diagnosis from the anomaly phenomenon to the root cause of the failure.

[0039] In one embodiment of the present invention, after determining the root causes of all performance anomalies within the current test cycle, a test analysis report is generated. This report records the spatiotemporal coordinates of each performance anomaly, the set of hypotheses for the anomaly's cause, the verification test process, and the final result of the root cause determination. This test analysis report is associated and stored with the corresponding original multi-source sensor data stream, the flushing valve's integrated state field, and the decoded internal operating state sequence to form a complete test case containing the problem, data, analysis, and conclusions. The complete test case is added to the historical case knowledge base. The historical case knowledge base is used to optimize the parameters of the improved Hidden Markov Model, enrich the causal links of the flushing valve fault knowledge graph, and update the expected response data under standard fault modes. When optimizing the parameters of the improved Hidden Markov Model, all complete test cases marked with determined root causes are selected from the historical case knowledge base. From each selected complete test case, its flushing valve integrated state field and internal operating state sequence are extracted. All extracted flushing valve integrated state fields are used as a new observation sequence training set, and all extracted internal operating state sequences are used as a new state sequence label set. A supervised learning algorithm is employed to re-estimate the observation probability matrix of the improved Hidden Markov Model using a new training set of observation sequences and a new set of state sequence labels. During the re-estimation process, anomalous observations that occur frequently in the root cause cases of failures are given higher probability weights in the observation probability matrix, thereby optimizing the model's ability to identify anomalous states.

[0040] In practical implementation, the system completes the knowledge accumulation and model optimization process after testing. After determining the root causes of all performance anomalies within the current testing cycle, the system automatically generates a structured test analysis report. This report, in electronic document format, records the spatiotemporal coordinates of each identified performance anomaly, the set of anomaly cause hypotheses generated through reverse tracing, the targeted verification test plan executed to validate these hypotheses, the collected high-precision response data, and the final root cause determination result. The test analysis report is associated and stored with the corresponding original multi-source sensor data stream, the flushing valve integrated state field generated through fusion processing, and the internal working state sequence decoded through an improved Hidden Markov Model. This association is achieved by assigning a unique identifier to each test case and using this identifier in the database to link all the aforementioned data and documents, thus forming a complete test case containing raw data, intermediate processing results, analysis process, and final conclusions.

[0041] In practice, complete test cases are added to the historical case knowledge base. This historical case knowledge base is a structured database that stores all complete test cases for which diagnoses have been completed. It is used to optimize the parameters of the improved Hidden Markov Model, enrich the causal links in the flushing valve fault knowledge graph, and update the expected response data under standard fault modes. For example, when a new "seal ring incision leakage" fault is confirmed, its corresponding complete test case is stored in the historical case knowledge base. The abnormal observation patterns extracted from this case can be used to match similar anomalies in the future. The correlation between the confirmed fault cause and the observed phenomenon can be added as a new causal link to the flushing valve fault knowledge graph. The high-precision acoustic emission signals collected during the verification test phase can be updated as expected response data for this fault mode.

[0042] In some embodiments, all complete test cases labeled with identified root causes of failures are selected from a historical case knowledge base. From each selected complete test case, its flushing valve integrated state field and internal operating state sequence are extracted. The flushing valve integrated state field is used as the observation sequence, and the internal operating state sequence is used as the corresponding true state label. All extracted flushing valve integrated state fields are used as a new observation sequence training set, and all extracted internal operating state sequences are used as a new state sequence label set. A supervised learning algorithm is used to re-estimate the observation probability matrix of the improved Hidden Markov Model using the new observation sequence training set and the new state sequence label set. The re-estimation of the observation probability matrix can be achieved by maximizing the likelihood function of the training data. The re-estimation formula for each element in the observation probability matrix can be expressed as: , in: This represents the probability of observing a discrete symbol in the hidden state after re-estimation. It is the time length of a single observation sequence. It is the number of complete test cases used for training in the historical case knowledge base. This represents the probability of being in state at time i of the training case (calculated using the forward-backward algorithm). It is the observation value of the i-th case at time t (i.e., the discretized representation of the overall state vector of the flushing valve). It is an indicator function; it takes the value 1 when the condition is met, and 0 otherwise.

[0043] Optionally, during the re-estimation process, anomalous observations that occur frequently in the root cause cases are given higher probability weights in the observation probability matrix, thereby optimizing the model's ability to identify anomalous states. Specifically, when calculating the weighted sum, for observations from root cause cases labeled "valve core wear," when the observation falls within the wear-related feature range, it can be multiplied by a weighting factor greater than 1 before summing. (For example This improves the estimated probability of abnormal observations under abnormal conditions (such as abnormal valve closure). Understandably, in this way, the improved Hidden Markov Model (HMM) can learn from historical failure cases, making its observation probability distribution more reflective of the sensor data characteristics under real failures. In some embodiments, the historical case knowledge base can be updated and the model optimized periodically, for example, triggering a batch retraining of the model after accumulating 100 new complete test cases. Understandably, the continuous accumulation of complete test cases enriches the historical case knowledge base, and periodic optimization based on this knowledge base allows the improved HMM, the flushing valve failure knowledge graph, and the standard failure mode library to continuously evolve, thereby continuously enhancing the diagnostic accuracy and generalization ability of the entire testing system.

[0044] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent testing of flush valves based on multi-source sensory data, characterized in that, The method includes: Acquire multi-source sensor data streams of the flushing valve during the test cycle. The multi-source sensor data streams include dynamic waveforms of water pressure, flow pulse sequences, valve body vibration spectrum, and sealing surface image frames. The multi-source sensor data stream is spatiotemporally aligned and fused to generate a comprehensive state field of the flushing valve within the test cycle. Based on the comprehensive state field of the flushing valve, an improved hidden Markov model is used to decode the internal working state sequence of the flushing valve. The improved hidden Markov model constrains the state transition probability based on the mechanical action stage characteristics of the flushing valve. Based on the decoded internal working state sequence, locate the abnormal performance events of the flushing valve and the corresponding spatiotemporal coordinates of the events; Based on the performance anomaly events and their corresponding spatiotemporal coordinates, a set of hypotheses for the causes of the anomalies is generated by reverse tracing, and the set of hypotheses for the causes of the anomalies is verified and tested to determine the root cause of the fault.

2. The multi-sensor data based flush valve intelligence testing method of claim 1, wherein, The multi-source sensor data streams are spatiotemporally aligned and fused to generate a comprehensive state field of the flushing valve within the test cycle, including: The dynamic waveform of water pressure is subjected to a segmented stability test, which divides the pressure stabilization stage, the pressure rise stage, and the pressure fall stage. The peak value and pulse width are simultaneously measured on the flow pulse sequence to identify the start time, peak time, end time and pulse intensity of each flow pulse; Time-frequency analysis was performed on the vibration spectrum of the valve body to extract the energy change curves of the characteristic frequency bands corresponding to valve core impact, spring reset, and sealing surface friction over time. Perform time-series analysis on the image frames of the sealing surface, calculate the pixel gradient changes in the sealing area in each frame, and generate a sealing surface deformation sequence; A unified test time axis is established, which maps the divided pressure stabilization phase, pressure rise phase, and pressure fall phase, the start time, peak time, end time, and pulse intensity of the identified flow pulse, the energy change curve of the extracted characteristic frequency band over time, and the generated sealing surface deformation sequence to the unified test time axis. On the unified test time axis, for each time slice, the pressure stage label, flow pulse parameter, vibration energy value and deformation value mapped to this slice are tensor spliced ​​to form a comprehensive state vector of the time slice. The combined state vectors of all time slices are combined in chronological order to form the combined state field of the flushing valve.

3. The multi-sensor data based flush valve intelligence testing method of claim 2, wherein, The method of decoding the internal operating state sequence of the flushing valve using an improved hidden Markov model includes: The improved hidden Markov model contains a set of hidden states, which correspond one-to-one with the physical working stages of the flushing valve, including the water inlet pressurization state, valve open state, high-speed flushing state, valve closed state, and reset sealing state. The improved Hidden Markov Model constrains the state transition probability based on the mechanical action stage characteristics of the flushing valve. Specifically, the state transition is only allowed to proceed unidirectionally in the order of water inlet pressure storage state, valve open state, high-speed flushing state, valve closed state, and reset sealing state, or remain in the current state. The integrated state vector of each time slice in the integrated state field of the flushing valve is used as the observation of the improved hidden Markov model at the current time step. Using the Viterbi algorithm, we find the hidden state sequence under a given observation sequence, which is the decoded internal working state sequence; During the decoding process, the observation probability distribution of the water inlet pressure storage state and the valve closed state is weighted and adjusted based on the stage division results of the water pressure dynamic waveform.

4. The multi-sensor data based flush valve intelligence testing method of claim 3, wherein, Using the Viterbi algorithm, a hidden state sequence is found given an observation sequence. This hidden state sequence is the decoded internal working state sequence, including: The comprehensive state field of the flushing valve is arranged in the order of the test time axis to form an observation sequence, and the state transition probability matrix, observation probability matrix and initial state probability vector of the improved hidden Markov model are used as known parameters. During the initialization phase of the Viterbi algorithm, the initial probability of each hidden state corresponding to the current observation value at the first time slice is calculated, and the probability value of each state and the previous state being empty are recorded. In the recursive phase of the Viterbi algorithm, the observation sequence is traversed backward from the second time slice. For each hidden state in the current time slice, the probability of transitioning from each state in the previous time slice to the current state and generating the current observation value is calculated. The path with the highest probability is selected from all transition paths, and the maximum probability value and the corresponding previous state are recorded. In the termination phase of the Viterbi algorithm, after traversing all observations, the state with the highest probability is selected from all states of the last time slice as the endpoint of the optimal path. Starting from the end of the optimal path, based on the previous state information recorded in the recursive stage, backtrack to the first time slice, and the resulting hidden state sequence is the hidden state sequence under the given observation sequence. The hidden state sequence is output as the decoded internal working state sequence.

5. The multi-sensor data based flush valve intelligence testing method of claim 4, wherein, The step of locating the abnormal performance events of the flushing valve and their corresponding spatiotemporal coordinates based on the decoded internal working state sequence includes: The decoded internal working state sequence is compared with the standard working state sequence, which defines the standard duration range and standard transition relationship of each hidden state under normal working conditions. Identify state timeout events or state insufficiency events in the internal working state sequence whose state duration exceeds the standard duration range; Identify state transition events or state missing events in the internal working state sequence that violate the standard transition relationship; In the overall state field of the flushing valve, find the time slices corresponding to the state timeout event, state insufficiency event, state jump event, or state missing event. From the comprehensive state vector corresponding to the time slice found, extract the abnormal pressure stage label, flow pulse parameter, vibration energy value and deformation value, and encapsulate the extracted abnormal parameters and event type to form the performance abnormal event. The position of the performance anomaly event and its corresponding time slice on the test time axis are collectively defined as the spatiotemporal coordinates of the event.

6. The multi-sensor data based flush valve intelligence testing method of claim 5, wherein, Based on the aforementioned performance anomaly events and their corresponding spatiotemporal coordinates, a set of hypotheses for the causes of the anomalies is generated through reverse tracing, including: Establish a knowledge graph for flush valve failure, wherein the knowledge graph uses performance abnormality events as nodes, component failure causes as nodes, and causal links between events and causes as edges; The performance abnormality events are input into the flushing valve fault knowledge graph for matching and retrieval. Based on the matched causal links, the source of one or more candidate component failures leading to the performance anomaly event is traced, forming an initial set of anomaly cause hypotheses; Based on the spatiotemporal coordinates corresponding to the event, a multi-source sensor data subfield within the time window preceding the occurrence of the abnormal event is extracted from the comprehensive state field of the flushing valve. Analyze the changing trends and collaborative relationships of different sensor data in the multi-source sensor data subfield, and score the support of each candidate component failure cause in the initial set of anomaly cause hypotheses. The candidate component failure causes are sorted and filtered based on the support scores to generate a set of anomaly cause hypotheses.

7. The intelligent testing method for flushing valves based on multi-source sensor data according to claim 6, characterized in that, Verification tests are performed on the set of hypotheses for the causes of the anomalies to determine the root cause of the failure, including: For each candidate component failure cause in the set of anomaly cause hypotheses, a targeted verification test plan is designed. The targeted verification test plan includes applying a specific excitation signal, acquiring higher-precision local sensing data, or executing a specific action sequence. On the flushing valve test platform, each of the aforementioned targeted verification test schemes was executed in sequence, and high-precision response data was collected synchronously during the verification test. The collected high-precision response data is compared with the expected response data of the corresponding candidate component failure cause under the standard failure mode; Calculate the consistency metric for each comparison, which integrates waveform similarity, feature matching degree, and time sequence alignment degree; Candidate component failure causes whose consistency metric values ​​exceed a preset confidence threshold are identified as verified root causes of failure. If the consistency metric values ​​of all candidate component failure causes do not exceed the preset confidence threshold, then the performance anomaly event and the corresponding multi-source sensor data subfield are marked as unknown anomaly mode.

8. The intelligent testing method for flushing valves based on multi-source sensor data according to claim 7, characterized in that, The collected high-precision response data is compared with the expected response data of the corresponding candidate component failure causes under standard failure modes, including: For candidate component failure causes corresponding to valve core wear failure, extract the valve core motion acceleration waveform and sealing surface leakage acoustic emission signal from the high-precision response data; The extracted valve core motion acceleration waveform and the expected acceleration waveform under the standard valve core wear fault mode are dynamically time-warped, and the path distance between the valve core motion acceleration waveform and the standard valve core wear fault mode is calculated as the waveform difference degree. The extracted acoustic emission signal from the sealing surface leakage was subjected to spectral analysis, and the ratio between its dominant frequency component energy and the dominant frequency component energy of the standard leakage spectrum was calculated as the spectral fit. By combining the waveform difference and the spectrum matching, a consistency metric for valve core wear faults is obtained through weighted summation.

9. The intelligent testing method for flushing valves based on multi-source sensor data according to claim 7, characterized in that, The method further includes: After determining the root cause of all performance anomalies in the current test cycle, a test analysis report is generated. The test analysis report records the spatiotemporal coordinates of each performance anomaly, the set of anomaly cause hypotheses, the verification test process, and the final result of the root cause determination. The test analysis report is associated and stored with the corresponding multi-source sensor data stream, flushing valve integrated state field and internal working state sequence to form a complete test case; The complete test cases are added to the historical case knowledge base to optimize the parameters of the improved hidden Markov model, enrich the causal links of the flushing valve fault knowledge graph, and update the expected response data under standard fault modes.

10. The multi-sensor data based flush valve intelligence testing method of claim 9, wherein, The complete test cases are added to the historical test case knowledge base to optimize the parameters of the improved Hidden Markov Model, including: From the historical case knowledge base, select all complete test cases marked with a definite root cause of the failure; Extract the overall state field and internal working state sequence of the flushing valve from each selected complete test case; Use all extracted flush valve integrated state fields as the new observation sequence training set, and use all extracted internal working state sequences as the new state sequence label set; A supervised learning algorithm is used to re-estimate the observation probability matrix of the improved Hidden Markov Model using the new observation sequence training set and the new state sequence label set. During the re-estimation process, abnormal observations that occur frequently in the root cause cases of failure are given higher probability weights in the observation probability matrix, thereby optimizing the model's ability to identify abnormal states.

Images