Injection pump electrical fault diagnosis method based on current harmonic analysis

Abstract

The present application relates to the technical field of medical equipment fault diagnosis, and discloses an injection pump electrical fault diagnosis method based on current harmonic analysis. The method comprises obtaining current harmonic reference data under the motor health state through initialization calibration. In operation, the real-time current signal is normalized with the reference to generate a real-time harmonic spectrum graph highlighting the deviation. The characteristic vector is extracted and matched with the fault characteristic library to calculate the similarity index. Based on the index, a multi-level fault judgment process including a cross-validation step is started to determine the fault type and level and generate a disposal instruction sequence. While executing the instructions, the method continuously monitors the spectrum changes and feeds back to the judgment process to realize dynamic correction of the diagnosis result. The method improves the detection sensitivity of early weak fault characteristics through differential comparison, improves the accuracy and reliability of diagnosis by using a closed-loop verification mechanism, and reduces false positives and false negatives.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical equipment fault diagnosis, in particular to an injection pump electrical fault diagnosis method based on current harmonic analysis. BACKGROUND

[0002] As a key device for precise drug delivery, the running reliability of the injection pump is directly related to the safety and effectiveness of patient treatment. The electrical state of the motor is a core factor that determines the performance of the injection pump. Potential faults such as winding insulation aging, bearing wear, and load mutation will eventually be reflected in the changes of motor current characteristics. Monitoring and analyzing the motor current is the mainstream technical path to achieve early warning and diagnosis of injection pump faults.

[0003] Existing injection pump fault diagnosis technologies mostly rely on simple threshold monitoring of motor operating current. Another common solution is to collect current signals and perform fast Fourier transform to obtain frequency spectrum, and then observe the amplitude changes of specific harmonics. These methods have limitations. The current threshold method has a reaction lag and can only trigger an alarm when the fault has developed to a considerable extent and caused significant current abnormalities, which cannot achieve early warning. The analysis method for fixed harmonics is difficult to separate the effects of the device's inherent electromagnetic characteristics, mechanical assembly differences, and normal operating fluctuations, resulting in insufficient sensitivity and specificity of diagnosis, and easy false positives or false negatives.

[0004] The root cause of these technical defects lies in the fact that traditional methods are based on the absolute value of the current signal, while each injection pump motor is a unique individual, and the initial "healthy harmonic fingerprint" is different. Directly using a unified and fixed threshold standard to measure all devices cannot effectively identify the weak characteristic changes caused by early faults. Existing diagnosis logic is mostly a simple "condition-action" mode, lacking a deep verification mechanism for suspected faults. Once the diagnosis conclusion is formed, it is fixed and cannot be modified according to system feedback, making it difficult to cope with complex and dynamically evolving fault scenarios. SUMMARY

[0005] The purpose of the present application is to provide an injection pump electrical fault diagnosis method based on current harmonic analysis to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application provides an injection pump electrical fault diagnosis method based on current harmonic analysis, which comprises:

[0007] Before the injection pump is put into operation, an initialization calibration process is performed to obtain the initial current harmonic data of the injection pump motor;

[0008] During the operation of the injection pump, the motor operating current signal is continuously captured, and the captured operating current signal is normalized based on the initial current harmonic data to generate a real-time harmonic spectrum;

[0009] Multi-dimensional feature extraction is performed on the real-time harmonic spectrum to obtain a set of feature vectors representing the current electrical state of the motor;

[0010] The feature vectors are matched with reference feature vectors of a series of known fault modes pre-stored in a fault feature library to generate a set of similarity indexes;

[0011] According to the number and type of similarity indexes exceeding a preset threshold, a multi-level fault judgment process is started, which includes a cross-validation step for suspected faults;

[0012] According to the output results of the multi-level fault judgment process, the specific fault type and severity level are determined, and a fault handling instruction sequence with priority ranking is generated accordingly;

[0013] The instructions in the fault handling instruction sequence are executed, while monitoring the changes in the real-time harmonic spectrum, and the monitoring data is fed back to the multi-level fault judgment process for dynamic correction of the diagnosis results.

[0014] Preferably, the execution of the initialization calibration process to obtain the initial current harmonic data of the injection pump motor includes:

[0015] Under no load and various preset speeds of the injection pump, the motor current signal is collected to obtain a basic current sample set;

[0016] High-precision spectral analysis is performed on each current sample in the basic current sample set to extract the amplitude and phase of each harmonic component, forming a reference harmonic data set;

[0017] The relationship between harmonic amplitude and speed is analyzed to construct a harmonic amplitude-speed relationship surface;

[0018] The relationship between harmonic phase and speed is analyzed to construct a harmonic phase-speed relationship surface;

[0019] The harmonic amplitude-speed relationship surface and the harmonic phase-speed relationship surface are integrated to form the initial current harmonic data of the injection pump motor.

[0020] Preferably, the normalization of the captured operating current signal based on the initial current harmonic data to generate a real-time harmonic spectrum includes:

[0021] The real-time speed value of the current injection pump motor is identified;

[0022] According to the real-time speed value, a corresponding reference harmonic amplitude and a reference harmonic phase are queried from the initial current harmonic data of the injection pump motor;

[0023] Actual harmonic amplitudes and actual harmonic phases are parsed from the captured working current signal;

[0024] A ratio of the actual harmonic amplitudes to the reference harmonic amplitudes is calculated to obtain normalized harmonic amplitudes;

[0025] A difference between the actual harmonic phases and the reference harmonic phases is calculated to obtain normalized harmonic phases;

[0026] The real-time harmonic spectrum is composed of all the normalized harmonic amplitudes and the normalized harmonic phases.

[0027] Preferably, multi-dimensional feature extraction is performed on the real-time harmonic spectrum to obtain a group of feature vectors representing the current electrical state of the motor, including:

[0028] The feature vectors at least include harmonic energy distribution features and harmonic phase relationship features;

[0029] Normalized amplitudes of specific harmonic waves are extracted from the real-time harmonic spectrum to form a harmonic amplitude sub-vector;

[0030] Gradients between adjacent harmonic amplitudes in the harmonic amplitude sub-vector are calculated to form a harmonic amplitude change gradient sequence;

[0031] Normalized phase differences between specific harmonic waves are extracted from the real-time harmonic spectrum to form a harmonic phase relationship sub-vector;

[0032] A phase stability index is formed by calculating the fluctuation variance of the harmonic phase relationship sub-vector on the time axis;

[0033] The harmonic amplitude sub-vector, the harmonic amplitude change gradient sequence, the harmonic phase relationship sub-vector, and the phase stability index are combined to form the feature vector.

[0034] Preferably, the feature vector is matched with reference feature vectors of a series of known fault modes pre-stored in a fault feature library to generate a group of similarity indexes, including:

[0035] Each element in the feature vector is assigned a weight coefficient, which is determined based on the sensitivity of the element to a specific fault;

[0036] Weighted Euclidean distances between corresponding elements of the feature vector and each reference feature vector in the fault feature library are sequentially calculated;

[0037] mapping the calculated weighted Euclidean distance to a value between zero and one, to generate the similarity index for each known failure mode;

[0038] sorting all the similarity indices by value, and selecting the similarity indices with values higher than a preliminary screening threshold to form a candidate failure set.

[0039] Preferably, according to the number and type of similarity indices exceeding a preset threshold, a multi-stage failure judgment process is initiated, which includes a cross-validation step for suspected failures, comprising:

[0040] When there is only one similarity index significantly higher than the others in the candidate failure set, the failure mode corresponding to this similarity index is determined as the primary diagnosis result;

[0041] When there are multiple similarity indices with values close to each other and all higher than a preset threshold in the candidate failure set, a cross-validation step is initiated: the occurrence frequency and co-occurrence probability of the multiple failure modes in the historical diagnosis records are queried, and the conditional probability of each candidate failure mode is calculated in combination with the working time of the current syringe pump;

[0042] The candidate failure mode with the highest conditional probability is selected as the dominant suspected failure, and the remaining candidate failure modes are selected as auxiliary suspected failures;

[0043] For the dominant suspected failure, the weight coefficient distribution strategy of the feature vector is adjusted to emphasize the weight of elements strongly related to this failure, and a second round of matching calculation is performed to generate a verification similarity index;

[0044] If the verification similarity index is significantly improved compared to the first round of calculation, the dominant suspected failure is confirmed as the final diagnosis result.

[0045] Preferably, according to the output results of the multi-stage failure judgment process, the specific failure type and severity level are determined, and a fault handling instruction sequence with priority ranking is generated accordingly, comprising:

[0046] According to the failure mode corresponding to the final diagnosis result, the corresponding basic handling instruction set is found from the handling strategy mapping table;

[0047] According to the value of the similarity index in the final diagnosis result, the severity level of the failure is mapped;

[0048] According to the severity level, the parameters of the instructions in the basic handling instruction set are refined, such as reducing the specific percentage of motor power or triggering the level of alarm;

[0049] determining whether there are multiple faults or abnormal states to be handled, prioritizing all the handling instructions according to the influence degree of the fault mode on the safety of the device;

[0050] generating the sequence of fault handling instructions in the order of priority from high to low.

[0051] Preferably, the feedback of the monitoring data to the multi-stage fault determination process for dynamic correction of the diagnostic result comprises:

[0052] executing the instruction with the highest priority in the sequence of fault handling instructions;

[0053] continuously obtaining a new real-time harmonic spectrum within a preset time after executing the instruction;

[0054] extracting a key feature value related to the confirmed fault from the new real-time harmonic spectrum and observing its trend;

[0055] if the key feature value returns to the normal range, continue to execute the handling instruction with the next priority;

[0056] if the key feature value does not improve or further deteriorates, suspend the subsequent instruction and trigger a diagnostic backtracking mechanism: recalculate the similarity of the current feature vector and the fault feature library, check whether there is a misdiagnosis or a new fault;

[0057] update the sequence of fault handling instructions according to the result of the diagnostic backtracking mechanism.

[0058] Preferably, the method further comprises an online self-learning update process of the fault feature library, which comprises:

[0059] record the data of each complete diagnostic process, including the final confirmed feature vector, the diagnostic result and the effect data after handling;

[0060] periodically perform cluster analysis on the historical records to identify new feature vector clusters that are not defined in the existing fault feature library;

[0061] for stable new feature vector clusters, confirm their corresponding physical fault modes by maintenance personnel combined with the on-site situation, and add them as new reference feature vectors to the fault feature library;

[0062] for existing reference feature vectors, fine-tune their vector values according to their performance in new diagnostic cases to cover the normal fluctuation range of the fault mode.

[0063] Preferably, the method further comprises a recalibration process after the overhaul or replacement of key components of the injection pump, which comprises:

[0064] Trigger the same step as the initialization calibration process to establish the new injection pump motor initial current harmonic data;

[0065] Compare the new injection pump motor initial current harmonic data with the old profile to calculate the profile difference;

[0066] If the profile difference exceeds the allowed range, prompt the fault feature library to perform overall offset correction;

[0067] Batch transform the reference feature vectors in the old fault feature library based on the profile difference to adapt to the new electrical characteristics of the injection pump, and form an updated fault feature library.

[0068] Compared with the prior art, the beneficial effects of the present application are:

[0069] By performing initialization calibration before the injection pump is put into operation, the initial current harmonic data of the motor in the healthy state is obtained as a dedicated reference. In subsequent operation, the real-time captured working current signal is normalized with this reference. The core of the analysis is transferred from the absolute amplitude of the current harmonic to the relative change amount relative to the healthy reference. The individual differences of the motor, inherent electromagnetic noise and constant load background effectively eliminate the basic influence of the harmonic features, so that the weak harmonic components caused by early faults superimposed on the normal signal are highlighted. The real-time harmonic spectrum generated by the normalization process improves the signal-to-noise ratio of the fault features.

[0070] A multi-level fault judgment process including a cross-validation step is introduced, and the feedback change of the harmonic spectrum is continuously monitored after the output treatment instruction. When the matching result of the feature vector and the fault library exceeds the threshold value, the system does not immediately give the final diagnosis, but starts the multi-level logic to cross-verify and confirm the specificity of the fault mode from multiple feature dimensions, and exclude similar interference. The system continuously monitors the change trend of the real-time harmonic spectrum while executing the treatment instruction, and feeds back this data to the judgment process. This forms a closed-loop diagnosis-action-verification-correction loop. The effect of the treatment measures is used as a key basis for verifying the correctness of the initial diagnosis, and drives the dynamic adjustment and optimization of the diagnosis conclusion and subsequent instructions. BRIEF DESCRIPTION OF DRAWINGS

[0071] Figure 1 The working principle diagram of the injection pump electrical fault diagnosis method based on current harmonic analysis described in the present application;

[0072] Figure 2 The flowchart for obtaining initial current harmonic data for the initialization calibration process;

[0073] Figure 3 The flowchart for obtaining feature vectors for multi-dimensional feature extraction;

[0074] Figure 4 Harmonic energy distribution characteristic fluctuation chart for injection pump fault interval

[0075] Figure 5 Comparison chart of current harmonic characteristic vector for injection pump electrical fault mode DETAILED DESCRIPTION

[0076] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0077] Please refer to Figure 1 The present application provides an injection pump electrical fault diagnosis method based on current harmonic analysis. The method realizes early diagnosis and processing of electrical faults by collecting and analyzing current harmonic data of the motor of the injection pump. The overall implementation scheme of the method is as follows:

[0078] Before the injection pump is put into operation, an initialization calibration process is performed. The process collects motor current signals under no load and at various preset speeds, and performs high-precision spectrum analysis to obtain initial current harmonic data of the motor of the injection pump, including a harmonic amplitude-speed relationship surface and a harmonic phase-speed relationship surface. During the operation of the injection pump, the working current signal of the motor is continuously captured, the reference harmonic amplitude and the reference harmonic phase are queried from the initial current harmonic data according to the real-time speed value, and then the working current signal is analyzed to calculate the normalized harmonic amplitude by calculating the ratio of the actual harmonic amplitude to the reference harmonic amplitude, and to calculate the normalized harmonic phase by calculating the difference between the actual harmonic phase and the reference harmonic phase, thereby generating a real-time harmonic spectrum chart. Multi-dimensional feature extraction is performed on the real-time harmonic spectrum chart to extract harmonic energy distribution features and harmonic phase relationship features, including harmonic amplitude sub-vectors, harmonic amplitude change gradient sequences, harmonic phase relationship sub-vectors and phase stability indicators, which are combined to form a feature vector. The feature vector is matched with the reference feature vectors in the fault feature library item by item, the similarity index is generated by weighted Euclidean distance and mapping, and the candidate fault set is screened. According to the number and type of the similarity index exceeding the preset threshold, a multi-level fault judgment process is started, which includes a primary diagnosis and a cross-validation step, and the fault mode is confirmed by calculating the conditional probability and adjusting the weight coefficient. According to the output result of the multi-level fault judgment process, the fault type and severity level are determined from the disposition strategy mapping table, and a priority-ordered fault disposition instruction sequence is generated. Finally, the instruction sequence is executed, the change of the real-time harmonic spectrum chart is monitored, and the feedback data is used for dynamic correction of the diagnosis result.

[0079] Example 1: See Figure 2 In the initial calibration process, the syringe pump operates at various preset speeds under no-load conditions, collecting motor current signals to form a basic current sample set. High-precision spectrum analysis is performed on each current sample to extract the amplitude and phase of each harmonic component, forming a reference harmonic dataset. The variation of harmonic amplitude with speed is analyzed to construct a harmonic amplitude-speed relationship surface, and the variation of harmonic phase with speed is analyzed to construct a harmonic phase-speed relationship surface. These two surfaces are integrated to form the initial current harmonic data. When generating the real-time harmonic spectrum, the real-time speed of the syringe pump motor is identified. Based on the real-time speed value, the corresponding reference harmonic amplitude and phase are retrieved from the initial current harmonic data. The actual harmonic amplitude and phase are parsed from the captured operating current signal. The ratio of the actual harmonic amplitude to the reference harmonic amplitude is calculated to obtain the normalized harmonic amplitude, and the difference between the actual harmonic phase and the reference harmonic phase is calculated to obtain the normalized harmonic phase. All normalized harmonic amplitudes and phases constitute the real-time harmonic spectrum.

[0080] In practice, the initial calibration process is performed before the injection pump is put into operation. The injection pump operates under no-load conditions at various preset speeds. The range of preset speeds should cover the entire effective operating range of the injection pump motor, for example, from a minimum speed of 100 rpm to a maximum speed of 1000 rpm, with speed points set at 50 rpm intervals. A high-precision current sensor is used to collect the motor current signal, obtaining a basic current sample set. This basic current sample set contains current time-series data at multiple speed points. High-precision spectral analysis is performed on each current sample in the basic current sample set. Specifically, the high-precision spectral analysis can be implemented using the Fast Fourier Transform (FFT) algorithm. The sampling frequency should be at least 2.56 times the highest harmonic frequency component of the motor current, for example, set to 10 kHz. A Hanning window can be selected as the window function to reduce spectral leakage, and the spectral resolution should be set sufficient to separate adjacent harmonic components, for example, set to 1 Hz. The FFT algorithm is used to extract the amplitude and phase of each harmonic component, forming a reference harmonic dataset. The reference harmonic dataset stores the amplitude vector and phase vector of each harmonic at each speed. The variation of harmonic amplitude with rotational speed in the reference harmonic dataset is analyzed, and a harmonic amplitude-rotational speed relationship surface is constructed. This surface maps discrete rotational speed-harmonic amplitude points to a continuous surface using a two-dimensional interpolation method. The interpolation formula can be expressed as:

[0081]

[0082] in: It is the first Subharmonics at rotational speed The interpolation magnitude below, It is the harmonic amplitude of the i-th sample point. These are interpolation weighting coefficients. is the number of sample points. The rule of variation of harmonic phase with rotational speed in the benchmark harmonic data set is analyzed, a harmonic phase-rotational speed relationship surface is constructed, the harmonic amplitude-rotational speed relationship surface and the harmonic phase-rotational speed relationship surface are integrated by using an interpolation method to process the phase data, and initial current harmonic data of the injection pump motor is formed. The initial current harmonic data is stored in the non-volatile memory in the form of a lookup table or a function.

[0083] In a specific implementation, the process of generating the real-time harmonic spectrum is continuously performed during the operation of the injection pump. The motor operating current signal is captured in real time by the embedded system, and the real-time rotational speed value of the current injection pump motor is identified by the encoder or the rotational speed sensor. According to the real-time rotational speed value, the corresponding benchmark harmonic amplitude and benchmark harmonic phase are queried from the initial current harmonic data of the injection pump motor. The query operation is based on interpolation calculation or direct lookup table access. The actual harmonic amplitude and the actual harmonic phase are parsed from the captured operating current signal. The parsing process uses a real-time spectrum analysis algorithm. The ratio of the actual harmonic amplitude to the corresponding benchmark harmonic amplitude is calculated to obtain the normalized harmonic amplitude. The difference between the actual harmonic phase and the corresponding benchmark harmonic phase is calculated to obtain the normalized harmonic phase. All the normalized harmonic amplitudes and the normalized harmonic phases are sorted by harmonic order to form a real-time harmonic spectrum. The real-time harmonic spectrum is stored in the memory in the form of an array or a matrix for subsequent processing. In some embodiments, the normalization process can include a moving average filter to suppress noise and ensure data stability. It can be understood that the generation frequency of the real-time harmonic spectrum is synchronized with the current sampling rate to adapt to dynamic working conditions. Optionally, the benchmark harmonic phase query can consider phase wrapping processing to ensure that the difference calculation is within a reasonable range. In some embodiments, the storage format of the initial current harmonic data is optimized to a block structure to improve query efficiency. It can be understood that the calculation of the normalized harmonic amplitude and the normalized harmonic phase is a scalar operation, which is easy to implement on a microcontroller. Optionally, the identification of the real-time rotational speed value can integrate multiple sensor data to improve reliability.

[0084] Embodiment 2: see Figure 3, the multi-dimensional feature extraction is performed from the real-time harmonic spectrogram, the feature vector contains harmonic energy distribution features and inter-harmonic phase relationship features, a normalized amplitude of a specific harmonic is extracted to form a harmonic amplitude sub-vector, a gradient between adjacent harmonic amplitudes in the harmonic amplitude sub-vector is calculated to form a harmonic amplitude variation gradient sequence, a normalized phase difference between specific harmonics is extracted to form a harmonic phase relationship sub-vector, a fluctuation variance of the harmonic phase relationship sub-vector on a time axis is calculated to form a phase stability index, and the harmonic amplitude sub-vector, the harmonic amplitude variation gradient sequence, the harmonic phase relationship sub-vector, and the phase stability index are combined to form the feature vector. In the matching calculation, a weight coefficient is assigned to each element in the feature vector based on the sensitivity of the element to a specific fault, the weighted Euclidean distance between the elements corresponding to the feature vector and each reference feature vector in the fault feature library is calculated, the weighted Euclidean distance is mapped to a value between zero and one to generate a similarity index, all similarity indexes are sorted and indexes higher than a preliminary screening threshold are selected to form a candidate fault set.

[0085] In specific implementations, the multi-dimensional feature extraction operation acts on the real-time harmonic spectrogram, the feature vector contains at least harmonic energy distribution features and inter-harmonic phase relationship features, a normalized amplitude of a specific harmonic is extracted from the real-time harmonic spectrogram, the specific harmonic usually includes the fundamental wave, the third harmonic, the fifth harmonic, and the seventh harmonic, these normalized amplitudes form a harmonic amplitude sub-vector, and the harmonic amplitude sub-vector is a one-dimensional array. The gradient between adjacent harmonic amplitudes in the harmonic amplitude sub-vector is calculated to form a harmonic amplitude variation gradient sequence, and the gradient calculation uses the backward difference method, for example, the difference between the normalized amplitudes of the fifth harmonic and the third harmonic. A normalized phase difference between specific harmonics is extracted from the real-time harmonic spectrogram, for example, the phase difference between the fundamental wave and the third harmonic, the phase difference between the third harmonic and the fifth harmonic, to form a harmonic phase relationship sub-vector, and the harmonic phase relationship sub-vector is another one-dimensional array. The fluctuation variance of the harmonic phase relationship sub-vector on the time axis is calculated to form a phase stability index, and the variance calculation is based on the data of multiple consecutive sampling points in a sliding time window. The harmonic amplitude sub-vector, the harmonic amplitude variation gradient sequence, the harmonic phase relationship sub-vector, and the phase stability index are combined to form the final feature vector, and the feature vector is a comprehensive numerical array. In some embodiments, the harmonic amplitude variation gradient sequence can contain higher-order differences to capture more complex variation patterns. It can be understood that the calculation of the phase stability index requires the caching of historical harmonic phase relationship sub-vector data.

[0086] In specific implementations, when matching the feature vector with the fault feature library, a weight coefficient is assigned to each element in the feature vector, the weight coefficient is determined based on the sensitivity of the element to a specific fault, and the sensitivity is obtained through historical fault data statistical analysis. The weighted Euclidean distance between the elements corresponding to the feature vector and each reference feature vector in the fault feature library is calculated in sequence, the weighted Euclidean distance The calculation formula is:

[0087]

[0088] wherein: is the dimension of the feature vector, is the jth element of the current feature vector, is the jth element of the reference feature vector, is the weight coefficient corresponding to the jth element. The calculated weighted Euclidean distance is mapped to a value between zero and one through a linear or nonlinear function to generate a similarity index for each known fault mode, and the mapping function ensures that the smaller the distance, the closer the similarity index is to one. All similarity indexes are sorted in descending order of value, and similarity indexes with values higher than a preliminary screening threshold are selected to form a candidate fault set, which contains fault mode identifiers and corresponding similarity indexes. In some embodiments, the weight coefficient distribution can be dynamically adjusted based on the feature contribution matrix of the fault mode. It can be understood that the weighted Euclidean distance calculation is the core operation step of the matching process. Optionally, the mapping function can use a Gaussian kernel function to enhance the discrimination. Optionally, the preliminary screening threshold can be configured according to the diagnostic strictness requirement.

[0089] Embodiment 3: The multi-level fault judgment process is started according to the similarity index. When there is only one similarity index in the candidate fault set that is significantly higher than the others, the fault mode corresponding to the index is determined as the primary diagnostic result. When there are multiple similarity indexes in the candidate fault set that are close in value and are all higher than a preset threshold, a cross-validation step is started, the occurrence frequency and co-occurrence probability of multiple fault modes in the historical diagnosis records are queried, the conditional probability of each candidate fault mode is calculated combining the working time of the current syringe pump, the candidate fault mode with the highest conditional probability is selected as the dominant suspected fault, and the rest are auxiliary suspected faults. The weight coefficient distribution strategy of the feature vector is adjusted for the dominant suspected fault to amplify the weight of the elements strongly related to the fault, and a second round of matching calculation is performed to generate a verification similarity index. If the verification similarity index is significantly improved compared to the first round of calculation value, the dominant suspected fault is confirmed as the final diagnostic result.

[0090] In specific implementations, the multi-stage fault judgment process is started according to the number and type of similarity indexes exceeding the preset threshold. When there is only one similarity index significantly higher than other indexes in the candidate fault set, the significance judgment is based on the difference between the similarity index and the second highest similarity index being greater than a set threshold, and then the fault mode corresponding to the similarity index is determined as the primary diagnosis result, and the process directly outputs this result. When there are multiple similarity indexes with close values and all higher than the preset threshold in the candidate fault set, the definition of close values is that the difference between the highest similarity index and the subsequent several similarity indexes is less than a specific tolerance, and then the cross-validation step is started. The cross-validation step first queries the occurrence frequency and co-occurrence probability of multiple fault modes in the historical diagnosis record. The occurrence frequency is the statistical frequency of each fault mode in the historical record, and the co-occurrence probability is the joint probability of multiple fault modes appearing at the same time in a diagnosis case. Combined with the working time length of the current injection pump, the working time length is read from the device running timer, the conditional probability of each candidate fault mode is calculated, and the conditional probability calculation is based on Bayes theorem. The candidate fault mode with the highest conditional probability is selected as the dominant suspected fault, and the remaining candidate fault modes are selected as auxiliary suspected faults, and the dominant suspected fault is taken as the key verification object.

[0091] In specific implementations, for the dominant suspected fault, the weight coefficient distribution strategy of the feature vector is adjusted to emphasize the weight of the strongly related elements to the fault, the strongly related elements are identified through the pre-defined feature-fault association table in the fault feature library, and the weight amplification coefficient is determined according to the association strength. The second round of matching calculation is performed to generate the verification similarity index, and the second round of matching calculation uses the adjusted weight coefficient to recalculate the weighted Euclidean distance and map it to the similarity. If the verification similarity index is significantly improved compared to the first round of calculation, the significant improvement refers to the growth of the similarity index exceeding the confirmation threshold, then the dominant suspected fault is confirmed as the final diagnosis result, and the process outputs this confirmation result. If the verification similarity index does not significantly improve or even decreases, then the dominant suspected fault is abandoned, and the candidate fault mode with the highest conditional probability is selected from the auxiliary suspected faults as a new dominant suspected fault, and the weight adjustment and verification calculation are repeated until a fault mode passes the verification or all candidate modes are excluded. In some embodiments, the calculation formula of the conditional probability can be expressed as: wherein: is the conditional probability of the fault mode occurring under the current context C, is the probability of the current context appearing when the fault mode occurs, is the prior probability of the fault mode , is the number of candidate fault modes. It is understood that the historical diagnosis records are stored in non-volatile memory and updated periodically. Optionally, the threshold and specific tolerance can be adaptively adjusted according to the diagnosis accuracy requirement. In some embodiments, the list of auxiliary suspected faults is dynamically maintained during the verification process. It is understood that the adjustment of the weight coefficient allocation strategy is the core operation of the cross-validation step. Optionally, when all candidate modes are excluded, the process can output an "unknown fault" result and trigger a more extensive detection mechanism.

[0092] In some embodiments, the output result of the multi-level fault judgment process is used to find the corresponding basic treatment instruction set of the fault mode from the treatment strategy mapping table, the severity level of the fault is mapped according to the numerical value of the similarity index, the parameters of the instructions in the basic treatment instruction set are refined according to the severity level, such as the specific percentage of reducing motor power or the level of triggering an alarm, it is determined whether there are multiple faults or abnormal states that need to be handled, all treatment instructions are prioritized according to the impact of the fault mode on the safety of the device, and a fault treatment instruction sequence is generated in order of priority from high to low. When executing the instruction sequence, the instruction with the highest priority is executed, and a new real-time harmonic spectrum is continuously obtained within a preset time, the key feature values related to the confirmed fault are extracted to observe their change trend, if the key feature values return to the normal range, the subsequent instructions are continued to be executed, if the key feature values do not improve or worsen, the subsequent instructions are paused and a diagnosis backtracking mechanism is triggered, the similarity between the current feature vector and the fault feature library is recalculated to check for misdiagnosis or new faults, and the fault treatment instruction sequence is updated according to the result of the diagnosis backtracking mechanism.

[0093] In specific implementation, the specific fault type and severity level are determined according to the output result of the multi-level fault judgment process, the output result of the multi-level fault judgment process includes the final diagnosis result and the corresponding similarity index value, the basic treatment instruction set matching the fault mode corresponding to the final diagnosis result is found from the treatment strategy mapping table, and Table 1 is referred to. The treatment strategy mapping table is a pre-defined database table.

[0094] Table 1: Treatment strategy mapping table

[0095] Failure mode identification Basic treatment instruction set description Typical parameter ranges F001 Reduce motor power; trigger alarm Power reduction 10-50%; alarm level 1-3 F002 Stop injection pump; send maintenance notification Immediate stop; high notification priority F003 Adjust drive frequency; log Frequency offset ±5%; log detail level 2

[0096] The severity level of the fault is mapped according to the numerical value of the similarity index in the final diagnosis result, the higher the numerical value of the similarity index, the greater the possibility of the fault but the lower the severity level, the mapping relationship adopts a piecewise linear function, and the calculation formula of the severity level L is: wherein: L is the severity level (1 is the least severe and 5 is the most severe), is the numerical value of the similarity index (between 0 and 1), represents rounding up. The instructions in the base set of handling instructions are parameterized according to the severity level, for example for the motor power reduction instruction, the power is reduced by 10% for severity level 1 and 50% for severity level 5. It is determined whether there are multiple faults or abnormal conditions to be handled, all handling instructions are prioritized according to the impact of the fault mode on the safety of the equipment, the impact is predefined based on the severity of the consequences that the fault mode can cause, and a sequence of fault handling instructions is generated by combining the instructions in order of priority from high to low, the sequence of fault handling instructions is an ordered list of instructions.

[0097] In specific implementations, the instructions in the sequence of fault handling instructions are executed and feedback is monitored, the instruction with the highest priority in the sequence of fault handling instructions is executed first, and a new real-time harmonic spectrum is obtained continuously for a preset time after the execution of the instruction, the preset time is set according to the type of the instruction, for example, 30 seconds after the execution of the power reduction instruction. The change trend of the key feature value related to the confirmed fault is observed by extracting the key feature value from the new real-time harmonic spectrum, the key feature value can be the normalized amplitude of a specific harmonic or a phase stability indicator. If the key feature value returns to the normal range, the handling instruction with the next priority is executed, and the normal range is set based on historical normal data. If the key feature value does not improve or further deteriorates, the subsequent instructions are suspended and a diagnostic backtracking mechanism is triggered, the diagnostic backtracking mechanism re-computes the similarity between the current feature vector and the fault feature library to check whether there is a misdiagnosis or a new fault. The sequence of fault handling instructions is updated according to the result of the diagnostic backtracking mechanism, and the update can include adding new instructions, modifying instruction parameters, or reordering. In some embodiments, the severity level mapping function can adopt a nonlinear form to better match the actual risk. It can be understood that the handling strategy mapping table needs to be updated regularly to cover newly discovered fault modes. Optionally, the change trend judgment of the key feature value can be based on statistical tests within a sliding window. In some embodiments, the diagnostic backtracking mechanism can call a simplified version of the multi-level fault judgment process to improve response speed. It can be understood that the setting of the preset time needs to balance the timeliness of the response and the reliability of the monitoring. Optionally, the update operation of the sequence of fault handling instructions needs to be logged for auditing and analysis.

[0098] Referring to Figure 4In the presented injection pump electrical fault diagnosis scenario, the trend of harmonic energy distribution characteristic is one of the core representation indexes of motor electrical state. In the specific analysis, the blue curve represents the real-time collected and processed harmonic energy distribution characteristic value, and the numerical fluctuation corresponds to the change of electrical characteristics in the motor operation process; the red area represents the fault interval, which is determined by matching the real-time harmonic spectrum feature vector. In the fault interval, the harmonic energy distribution characteristic value presents a significant and violent fluctuation: its value rapidly jumps from about 0.10 to above 0.30 before the start of the fault interval, and continuously maintains a high fluctuation range of 0.25-0.40 within the interval. This phenomenon corresponds to the abnormal performance of the motor electrical state deviating from the normal benchmark, which is due to the distortion of the electrical parameters of the motor winding, drive circuit and other components in the fault mode, causing the current harmonic amplitude to deviate from the benchmark harmonic amplitude-speed relationship curve established in the initialization calibration phase. The harmonic energy distribution characteristic value after normalization processing also presents a large fluctuation. After the end of the fault interval, the harmonic energy distribution characteristic value falls back to the normal fluctuation interval of about 0.10, corresponding to the regression of the motor electrical state to the normal range after the execution of the fault disposal instruction. The characteristic fluctuation law of the graph can be used as the core observation basis of the "key characteristic value trend change" in fault diagnosis, to verify the effectiveness of the fault disposal instruction and the judgment condition of triggering the diagnosis backtracking mechanism.

[0099] Example 5: The method includes an online self-learning update process of the fault feature library, records the data of each complete diagnosis process, including the final confirmed feature vector, diagnosis result and effect data after disposal, periodically analyzes the historical records to identify new feature vector clusters not defined in the fault feature library, and adds new reference feature vectors to the fault feature library after confirming the corresponding physical fault mode based on the site situation. For the existing reference feature vectors, the vector values are fine-tuned according to their performance in new diagnosis cases to cover the normal fluctuation range of the fault mode. The method also includes a recalibration process after the overhaul or replacement of key components of the injection pump, which triggers the same steps as the initialization calibration process to establish new initial current harmonic data of the injection pump motor. The new initial current harmonic data is compared with the old data to calculate the contour difference, and if the contour difference exceeds the allowed range, the fault feature library may need to be corrected as a whole. The reference feature vectors in the old fault feature library are batch-converted based on the contour difference to form an updated fault feature library.

[0100] In specific implementations, the online self-learning update process runs continuously, recording each complete diagnosis process data, including the final confirmed feature vector, diagnosis result and post-treatment effect data, which are time-stamped and stored in a historical database. The historical records are periodically clustered for analysis. The clustering analysis uses a density clustering algorithm to identify feature vector clusters, thereby discovering new patterns not defined in the existing fault feature library. For the identified new feature vector clusters, the stability assessment is based on the distribution density and duration of the data points within the cluster. After the corresponding physical fault pattern is confirmed by maintenance personnel in combination with the on-site operation log, the cluster center vector is added to the fault feature library as a new reference feature vector. For existing reference feature vectors, the fine-tuning process adjusts the vector values based on the performance in new diagnosis cases using a sliding window average method to cover the normal fluctuation range of the fault pattern. In some embodiments, the frequency of clustering analysis can be adaptively adjusted to balance the computational overhead and real-time performance. It can be understood that the historical database uses a ring buffer structure to manage the storage space. Optionally, the confirmation process of new feature vector clusters can integrate remote expert review functions.

[0101] In specific implementations, the recalibration process is automatically triggered after major repair or replacement of key components of the syringe pump, performing the same steps as the initialization calibration process, including no-load multi-speed testing and high-precision spectral analysis, establishing new syringe pump motor initial current harmonic data, comparing the new initial current harmonic data with the old data, calculating the profile difference, which quantifies the overall deviation between the new and old data, with the formula: wherein: is the profile difference, is the total number of harmonic features, is the kth feature value of the old initial current harmonic data, is the kth feature value of the new initial current harmonic data, is the weight coefficient of the kth feature based on its importance, if the profile difference exceeds the allowed range, the system generates a prompt message to suggest overall offset correction of the fault feature library. The correction process performs batch linear transformation of the reference feature vectors in the old fault feature library based on the profile difference, with the transformation parameters determined by least squares fitting, forming the updated fault feature library. In some embodiments, the profile difference allowed range can be dynamically configured according to the device accuracy requirements. Optionally, the recalibration process can selectively perform partial calibration steps to improve efficiency. It can be understood that the profile difference calculation requires normalization preprocessing of the new and old data. Optionally, the batch transformation operation can retain historical versions to support rollback.

[0102] Referring to Figure 5The figure presents the variation of the feature vector with the harmonic order under different failure modes (including normal operation). Specifically, the figure takes the harmonic order as the horizontal axis and the feature vector value as the vertical axis, and compares the feature vector curves corresponding to five states: normal operation, bearing wear, coil aging, power fluctuation, and mechanical jam. The feature vector value of the mechanical jam mode is at a relatively high level under each harmonic order and shows a continuous downward trend. The feature vector values of the coil aging and bearing wear modes gradually decrease with the increase of the harmonic order, and the curves of the two modes intersect. The feature vector value of the power fluctuation mode first decreases, then increases, and then continuously decreases. The feature vector value of the normal operation mode always maintains in a low interval. These feature vectors are extracted from the multi-dimensional features of the real-time harmonic spectrum, including harmonic amplitude, phase relationship, and other core information. The differences can be calculated by weighted Euclidean distance to calculate the similarity, providing a quantitative basis for fault mode matching and diagnosis.

[0103] While embodiments of the present application have been shown and described, it is to be understood that the embodiments described are merely divergences, modifications, replacements and variations of the embodiments, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for diagnosing electrical faults in an injection pump based on current harmonic analysis, characterized in that, The method includes the following steps: Before the syringe pump is put into operation, an initialization calibration procedure is performed to obtain the initial current harmonic data of the syringe pump motor; During the operation of the injection pump, the motor operating current signal is continuously captured, and the captured operating current signal is normalized based on the initial current harmonic data to generate a real-time harmonic spectrum. Multi-dimensional feature extraction is performed on the real-time harmonic spectrum to obtain a set of feature vectors characterizing the current electrical state of the motor; The feature vector is matched item by item with a series of reference feature vectors of known fault modes pre-stored in the fault feature library to generate a set of similarity indices. Based on the number and type of similarity indicators exceeding a preset threshold, a multi-level fault judgment process is initiated, which includes a cross-validation step for suspected faults. Based on the output of the multi-level fault judgment process, the specific fault type and severity level are determined, and a fault handling instruction sequence with priority order is generated accordingly. The instructions in the fault handling instruction sequence are executed, while the changes in the real-time harmonic spectrum are monitored, and the monitoring data is fed back to the multi-level fault judgment process for dynamic correction of the diagnostic results. The initialization calibration process for obtaining initial current harmonic data of the syringe pump motor includes: Under no-load and various preset speeds of the injection pump, the motor current signal is collected to obtain a basic current sample set; High-precision spectrum analysis is performed on each current sample in the basic current sample set to extract the amplitude and phase of each harmonic component, forming a reference harmonic dataset; The variation of harmonic amplitude with rotational speed in the reference harmonic data is analyzed, and a surface relating harmonic amplitude to rotational speed is constructed. The variation of harmonic phase with rotational speed in the reference harmonic data is analyzed, and a harmonic phase-rotational speed relationship surface is constructed. By integrating the harmonic amplitude-speed relationship surface and the harmonic phase-speed relationship surface, the initial current harmonic data of the injection pump motor is formed. The step of normalizing the captured operating current signal based on the initial current harmonic data to generate a real-time harmonic spectrum includes: Identify the real-time rotational speed of the current injection pump motor; Based on the real-time rotational speed value, the corresponding reference harmonic amplitude and reference harmonic phase are retrieved from the initial current harmonic data of the injection pump motor. The actual harmonic amplitude and phase are extracted from the captured operating current signal. Calculate the ratio of the actual harmonic amplitude to the corresponding reference harmonic amplitude to obtain the normalized harmonic amplitude; The difference between the actual harmonic phase and the corresponding reference harmonic phase is calculated to obtain the normalized harmonic phase; The real-time harmonic spectrum is composed of all the normalized harmonic amplitudes and the normalized harmonic phases.

2. The method according to claim 1, characterized in that, The process of extracting multi-dimensional features from the real-time harmonic spectrum yields a set of feature vectors characterizing the current electrical state of the motor, including: The feature vector includes at least the harmonic energy distribution characteristics and the phase relationship characteristics between harmonics; The normalized amplitude of a specific harmonic is extracted from the real-time harmonic spectrum diagram to form a harmonic amplitude sub-vector; Calculate the gradient between adjacent harmonic amplitudes in the harmonic amplitude sub-vector to form a harmonic amplitude change gradient sequence; The normalized phase difference between specific harmonics is extracted from the real-time harmonic spectrum diagram to form a harmonic phase relationship sub-vector; Calculate the fluctuation variance of the harmonic phase relationship sub-vector on the time axis to form a phase stability index; The feature vector is formed by combining the harmonic amplitude sub-vector, the harmonic amplitude change gradient sequence, the harmonic phase relationship sub-vector, and the phase stability index.

3. The method according to claim 2, characterized in that, The step involves performing a step-by-step matching calculation between the feature vector and a series of reference feature vectors of known fault modes pre-stored in a fault feature database to generate a set of similarity indices, including: Each element in the feature vector is assigned a weight coefficient, which is determined based on the element's sensitivity to a specific fault. Calculate the weighted Euclidean distance between the feature vector and the corresponding element of each reference feature vector in the fault feature library in turn; The calculated weighted Euclidean distance is mapped to a value between zero and one to generate the similarity index for each known failure mode; All the similarity indicators are sorted by numerical value, and similarity indicators with values ​​higher than the initial screening threshold are selected to form a candidate fault set.

4. The method according to claim 3, characterized in that, Based on the number and type of similarity indicators exceeding a preset threshold, a multi-level fault judgment process is initiated. This multi-level fault judgment process includes a cross-validation step for suspected faults, including: When only one similarity index in the candidate fault set is significantly higher than the others, the fault mode corresponding to that similarity index is determined to be a preliminary diagnostic result. When multiple similarity index values ​​in the candidate fault set are close and all exceed a preset threshold, the cross-validation step is initiated: query the occurrence frequency and co-occurrence probability of multiple fault modes in historical diagnostic records, and calculate the conditional probability of each candidate fault mode in combination with the current working time of the injection pump. Select the candidate fault mode with the highest conditional probability as the dominant suspected fault, and use the remaining candidate fault modes as auxiliary suspected faults. For the suspected dominant fault, the weight coefficient allocation strategy of the feature vector is adjusted, focusing on increasing the weight of elements strongly related to the fault, and a second round of matching calculation is performed to generate a similarity index for verification. If the similarity index used for verification shows a significant improvement compared to the first round of calculations, then the dominant suspected fault is confirmed as the final diagnostic result.

5. The method according to claim 4, characterized in that, The step involves determining the specific fault type and severity level based on the output of the multi-level fault judgment process, and generating a priority-ordered sequence of fault handling instructions accordingly, including: Based on the fault mode corresponding to the final diagnosis result, the corresponding basic handling instruction set is found from the handling strategy mapping table; The magnitude of the similarity index in the final diagnostic results is mapped to the severity level of the fault. Based on the severity level, the instructions in the basic handling instruction set are refined with parameters, such as the specific percentage of motor power reduction or the level of alarm triggering; Determine if there are multiple faults or abnormal states that need to be handled, and prioritize all handling instructions according to the degree of impact of the fault mode on equipment safety; The fault handling instruction sequence is generated by combining instructions in descending order of priority.

6. The method according to claim 5, characterized in that, The step of feeding monitoring data back to the multi-level fault diagnosis process for dynamic correction of diagnostic results includes: Execute the instruction with the highest priority in the fault handling instruction sequence; Within a preset time period after the instruction is executed, new real-time harmonic spectrum diagrams are continuously acquired; Extract key feature values ​​related to confirmed faults from the new real-time harmonic spectrum and observe their changing trends; If the key feature value returns to the normal range, then the subsequent priority processing instructions will continue to be executed; If the key feature values ​​do not improve or worsen further, subsequent instructions are paused and a diagnostic backtracking mechanism is triggered: the similarity between the current feature vector and the fault feature library is recalculated to check for misdiagnosis or new faults. The fault handling instruction sequence is updated based on the results of the diagnostic backtracking mechanism.

7. The method according to claim 1, characterized in that, The method also includes an online self-learning update process for the fault feature database, which includes: Record data from each complete diagnostic process, including the final confirmed feature vector, diagnostic results, and post-treatment effect data; Periodically perform cluster analysis on historical records to identify new feature vector clusters that are not defined in the existing fault feature library; For a stable new feature vector cluster, maintenance personnel confirm its corresponding physical fault mode based on the on-site situation and add it as a new reference feature vector to the fault feature library. For existing reference feature vectors, their vector values ​​are fine-tuned based on their performance in new diagnostic cases to cover the normal fluctuation range of the fault mode.

8. The method according to claim 7, characterized in that, The method also includes a recalibration process after a major overhaul or replacement of critical components of the syringe pump, which includes: Trigger the same steps as the initial calibration process to establish new initial current harmonic data for the syringe pump motor; The initial current harmonic data of the new injection pump motor is compared with the old profile to calculate the profile difference. If the contour difference exceeds the allowable range, it indicates that the fault feature library may need to undergo overall offset correction. The reference feature vectors in the old fault feature library are batch transformed based on the contour difference degree to adapt them to the new electrical characteristics of the injection pump, thus forming an updated fault feature library.

Images