Gas cylinder filling method, early warning method, checking method, system and medium
By using voice analysis and multimodal data evaluation methods, the problem of traditional assessment methods being unable to evaluate the psychological quality and operational stability of operators has been solved. This has enabled multi-dimensional evaluation and real-time risk warning of gas cylinder filling operators, thereby improving the level of safety production management.
Patent Information
- Application Number
- CN202610343853.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-20
- Publication Date
- 2026-04-21
AI Technical Summary
Existing gas cylinder filling operator qualification certification assessments cannot effectively evaluate operators' psychological qualities and operational stability, cannot identify minor operational irregularities during the filling process, and cannot warn of potential risks. Traditional assessment environments differ greatly from actual work scenarios, making it impossible to assess operators' emergency response capabilities under pressure.
A multimodal assessment method based on voice analysis is adopted. By collecting the operator's acoustic data and gas cylinder movement posture data, combined with a deep learning model, the operator's knowledge proficiency, psychological quality and operational stability are evaluated, and an acoustic risk identification model is used for real-time risk warning.
It enables multi-dimensional assessment of operator capabilities and real-time risk warning, objectively quantifies the comprehensive capabilities of operators, provides data-driven safety management decision support, and improves the prediction and control level of safe production.
Smart Images

Figure CN121903471A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of industrial safety production, and particularly to a filling method, a warning method, an assessment method, a system and a medium for gas cylinders. Background Art
[0002] The filling of gas cylinders, especially the filling operations involving high-pressure, flammable, explosive, toxic and other high-risk media, is a high-risk link in industrial safety production. The existing safety guarantee system mainly relies on the formulation of standard operating procedures (SOPs), the regular maintenance and repair of equipment, and the qualification assessment of operators. However, this system has profound and irreparable defects in dealing with accidents caused by "unsafe states of people". The current operator qualification assessment focuses on practical examinations. Such examinations usually require candidates to complete a complete filling process according to SOPs in a standard environment. The examiner judges whether the candidate's operation complies with the specifications by visual observation. This mode has several major limitations. First, the evaluation dimension is single and superficial. The assessment can only verify whether the candidate remembers the operation process, that is, "knowing how to do it", but cannot deeply evaluate the fineness and smoothness in actual operation. For example, rough operations and minor impacts during the handling of gas cylinders and the docking of valves, although they do not immediately cause accidents, will cause stress accumulation in the equipment and pose long-term safety hazards, and these details are often ignored by subjective judgments. Second, and the most fatal defect, is that it is completely impossible to assess the psychological quality and emergency response ability of operators. The traditional assessment environment is single and stress-free, which is vastly different from the real operation scenario, especially the tense atmosphere during abnormal working conditions. A person who operates proficiently in the examination room is very likely to become rigid in thinking and deformed in operation due to tension, anxiety or panic when facing sudden pressure abnormalities, abnormal equipment noises or minor leaks in the real world, and is unable to correctly execute the emergency plan, thus turning a controllable small problem into a catastrophic accident. This "psychological stability" under pressure and the "reactivity" to uncertain factors are the key to determining the bottom line of safety production, but they have always been an absolute blind spot in the existing assessment system. Third, it is impossible to give early warnings for risks directly caused by human factors such as operators' inattentiveness and non-standard operations. Although the current behavior recognition technology based on computer vision can identify macroscopic behaviors such as whether a safety helmet is worn and whether a dangerous area is entered, it is difficult to effectively capture and analyze the fine operation quality during the filling process and the weak signals indicating risks (such as high-frequency gas leakage sounds). Summary of the Invention
[0003] In order to solve the technical problem that the existing operator qualification assessment cannot effectively evaluate the psychological quality and operation stability of operators, the present invention provides a filling method, a warning method, an assessment method, a system and a medium for gas cylinders.
[0004] This invention is achieved through the following technical solution: In a first aspect, the present invention provides a method for safety assessment of gas cylinder filling based on voice analysis, comprising: Acoustic data was collected from the operators' verbal descriptions of operating procedures during performance assessments. This acoustic data was compared to the standard operating procedure (SOP) and an accuracy analysis was performed. Based on the accuracy, a knowledge proficiency score was calculated. S pro .
[0005] Acoustic features are extracted from acoustic data; these features are then input into an emotion recognition model incorporating an attention mechanism to identify emotional states; and psychological quality scores are calculated based on the duration and cumulative intensity of different emotional states. S emo .
[0006] Collect motion posture data of the gas cylinders; based on the motion posture data, obtain the duration of excessive tilting. T tilt Number of impacts N ram Impact J angular velocity variance Var ( ω The quantitative indicator group includes [the following]; the operational stability score is calculated based on the quantitative indicator group. S sta .
[0007] Will S pro , S emo , S sta The weighted fusion is used to obtain the assessment score. S total .
[0008] A passing score is set based on the assessment requirements; if S total If the score is greater than or equal to the passing score, the operator is deemed to have passed the assessment, and based on... S pro , S emo , S sta Operators are given multi-dimensional capability profiles and specialized training.
[0009] Secondly, this invention also proposes a risk warning method for gas cylinder filling based on voice analysis, which performs acoustic risk warning during operational assessments using the safety assessment method for gas cylinder filling based on voice analysis described in the first aspect. The acoustic risk warning method includes the following steps: Acoustic data is collected during the gas cylinder filling process; the acoustic data is input into the acoustic risk identification model for identification, and the probability distribution of risk categories is obtained; the risk categories include: "normal", "leakage" and "overpressure".
[0010] The acoustic risk identification model's identification process is as follows: first, acoustic data is converted into a two-dimensional Mel spectrogram; then, local time-frequency features representing different acoustic events are extracted from the Mel spectrogram; the local time-frequency features are processed by LSTM to capture the evolution of local time-frequency features in the time dimension and long-distance contextual relationships; finally, the output of LSTM is mapped to the probability distribution of risk categories after passing through a fully connected layer and activation function.
[0011] And / or, all neural network models used in the gas cylinder filling safety assessment methods must be trained before use; the training methods for the acoustic risk identification model include: In a controlled experimental environment, various normal and abnormal filling conditions were simulated, and audio and physical parameters, including pressure, valve temperature, and real-time filling flow rate, were simultaneously acquired. Based on the changes in physical parameters, corresponding risk event labels were assigned to the synchronized audio to generate a labeled training dataset. The acoustic risk identification model was then trained using this labeled training dataset.
[0012] The risk event labels include: when the derivative of pressure with respect to time exceeds the preset maximum slope, the audio corresponding to the same time will be labeled as an overpressure event; when the derivative of valve temperature with respect to time is less than the preset threshold, the audio corresponding to the same time will be labeled as a leakage event; when the pressure and real-time filling flow rate show abnormal stagnation or fluctuation, the audio corresponding to the same time will be labeled as a blockage event.
[0013] Thirdly, this invention also proposes a voice analysis-based safety assessment system for gas cylinder filling, which utilizes the voice analysis-based safety assessment method for gas cylinder filling described in the first aspect. The gas cylinder filling safety assessment system includes: a data acquisition unit, a scoring unit, and a judgment unit. The data acquisition unit collects acoustic data from the operator's verbal description of the operating procedures during the assessment. The scoring unit compares the acoustic data with the standard operating procedures and performs accuracy analysis; based on the accuracy, a knowledge proficiency score is calculated. S pro ; Extract acoustic features from acoustic data; input the acoustic features into an emotion recognition model incorporating an attention mechanism for emotion state recognition; calculate psychological quality scores based on the duration and cumulative intensity of different emotion states. S emo Collect motion posture data of the gas cylinders; based on the motion posture data, obtain the duration of excessive tilting. T tilt Number of impacts Nram Impact J angular velocity variance Var ( ω The quantitative indicator group includes [the following]; the operational stability score is calculated based on the quantitative indicator group. S sta ;Will S pro , S emo , S sta The weighted fusion is used to obtain the assessment score. S total The judgment unit is used to set a passing score based on the assessment requirements; if S total If the score is greater than or equal to the passing score, the operator is deemed to have passed the assessment, and based on... S pro , S emo , S sta Operators are given multi-dimensional capability profiles and specialized training.
[0014] Fourthly, this invention proposes a safe gas cylinder filling method based on voice analysis, which uses the voice analysis-based gas cylinder filling safety assessment method from the first aspect to assess operators. After passing the assessment, operators are permitted to fill gas cylinders within the assessment period. While the operator is filling the gas cylinder, the voice analysis-based gas cylinder filling risk warning method from the second aspect is used for acoustic risk warning. When "leakage" or "overpressure" is detected, an alarm is immediately triggered to remind the user to stop the gas cylinder filling operation.
[0015] Fifthly, the present invention provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the steps of the voice analysis-based gas cylinder filling safety assessment method in the first aspect.
[0016] The beneficial effects of this invention are as follows: 1. This invention, through innovative multimodal assessment, transforms operator competence from "relying on intuition" to "relying on data." It not only provides the final result but also identifies specific weaknesses in each stage, effectively assessing operators' psychological qualities and emergency response capabilities. Furthermore, through trend analysis, it can proactively identify potential risks and changes in personnel status, providing robust data support for management activities such as personnel certification, task allocation, and training optimization.
[0017] 2. This invention utilizes multimodal sensor data acquisition, deep learning model and physical model fusion analysis technology to achieve a comprehensive assessment of the professional competence of gas cylinder filling operators and real-time risk warning of the filling process. It is especially suitable for specific scenarios with extremely high safety requirements, such as hydrogen energy, industrial specialty gases, liquefied natural gas and other filling fields. Attached Figure Description
[0018] Figure 1 This is a flowchart of a gas cylinder filling safety assessment method based on voice analysis in one embodiment; Figure 2 This is a flowchart of a gas cylinder safety filling method based on voice analysis in another embodiment. Detailed Implementation
[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] It should be noted that when a component is said to be "installed on" another component, it can be directly on the other component or it may be in a component that is centered on it. When a component is said to be "set on" another component, it can be directly set on the other component or it may also be in a component that is centered on it. When a component is said to be "fixed to" another component, it can be directly fixed to the other component or it may also be in a component that is centered on it.
[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.
[0022] This invention aims to address the core pain points of traditional safety assessments, which fail to effectively evaluate operators' psychological qualities and operational stability, and lack effective real-time risk prediction methods during the assessment process and even daily operations. To this end, it constructs a closed-loop system from "assessment" to "real-time early warning." Through innovative data acquisition, processing, and modeling methods, it achieves in-depth, dynamic, and proactive control over the "human-machine-environment" safety elements. Based on this, please refer to... Figure 1This embodiment provides a method for assessing the safety of gas cylinder filling based on voice analysis. This method mainly consists of two parts: multimodal operator competence assessment and real-time risk warning based on acoustic features. By combining these two core functions, an application closed loop from "static assessment" to "dynamic protection" is constructed. The assessment stage not only screens qualified personnel, but the massive amounts of data generated also continuously provide nourishment for the iterative optimization of the subsequent warning system. Furthermore, the real risk events recorded by the warning system in daily work can feed back into the assessment system, enriching its case library. This continuous improvement closed loop of "assessment-screening-on-job-protection-retraining" elevates safety management from passive "post-event traceability" and static "access assessment" to a new level of proactive "pre-event prediction" and dynamic "process intervention," providing strong technical support for achieving inherent safety in the industrial field. The following is a detailed introduction to these two core functions: I. Multimodal operator competency assessment.
[0023] During the assessment, operators are required to verbally describe the procedures while performing the operation assessment. Acoustic data from the operator's verbal description and the motion posture data of the gas cylinder being operated are simultaneously collected using non-invasive sensors. For example, the operator's verbal description of the procedures includes the following aspects: I. Pre-filling inspection items: 1. Scan the gas cylinder's electronic tag. 2. Whether it is a gas cylinder owned by the unit for filling. 3. Whether it is within the specified inspection period. 4. Whether the outlet thread type of the cylinder valve matches the specified thread type for the gas being filled (flammable left-hand thread, non-flammable right-hand thread). 5. Whether there is residual pressure in the gas cylinder. 6. Inspection of the outer surface of the gas cylinder. 7. Whether the safety accessories of the gas cylinder are complete and meet safety requirements. 8. For gas cylinders containing oxidizing gases, whether the cylinder body, valve, etc., are contaminated with grease or other flammable materials. II. Inspection items during filling: 1. When the filling pressure inside the cylinder reaches 10% of the filling amount, an empty cylinder cannot be added. 2. When the filling pressure inside the cylinder reaches 7MPa, check that the temperature of the outer wall of the gas cylinder is normal. 3. Once the filling pressure inside the cylinder reaches 10MPa, check if the cylinder valve seal is intact. 4. The filling time should not be less than 30 minutes. III. Post-filling inspection items: 1. Whether the cylinder pressure (filling volume) meets the requirements. 2. Whether the seal of the cylinder valve and cylinder opening connection is intact. 3. Inspection of the cylinder's appearance and temperature. 4. Whether the cylinder's safety accessories are complete. 5. Whether the cylinder's filling label and warning label are affixed. The acoustic data is fed into a dual-channel parallel analysis engine. The first channel is the knowledge proficiency assessment module, which uses automatic speech recognition technology to convert speech into text, and then uses a natural language processing model to perform a deep comparison between the converted text and the preset standard operating procedure (SOP) text. By analyzing keyword coverage, process sequence accuracy, and semantic completeness, a quantitative knowledge proficiency score is generated. S proThe second channel is the psychological assessment module. It performs affective analysis on the same acoustic data, extracting and quantifying key acoustic features such as fundamental frequency, speech rate, energy, jitter, and shimmer. Using a deep learning model based on attention mechanisms, it assesses the candidate's psychological state, such as tension, anxiety, or confidence, in real time and outputs a dynamic psychological assessment score. S emo Meanwhile, by collecting motion attitude data from the inertial measurement unit (IMU) integrated on the gas cylinder, mathematical and physical modeling methods are used to calculate indicators such as the number of impacts, impact intensity (Jerk), and duration of excessive tilting during the operation. This allows for a precise assessment of the candidate's action standardization and stability in key aspects such as handling and valve connection, generating an operational smoothness score. S sta Finally, in the assessment scenario, a weighted fusion algorithm is used to combine the scores of these three independent dimensions ( S pro , S emo , S sta This data is integrated into a comprehensive evaluation result, providing an objective and reliable data-driven decision-making basis for personnel selection, training, and onboarding.
[0024] Specifically, the non-invasive sensors include: an audio acquisition unit, a motion and attitude acquisition unit, and a physical state acquisition unit. The audio acquisition unit uses a high-fidelity microphone with ambient noise suppression capabilities to clearly capture operator speech and ambient sounds during assessments and daily operations. The motion and attitude acquisition unit uses a nine-axis inertial measurement unit (IMU) integrating a three-axis accelerometer, a three-axis gyroscope, and a three-axis magnetometer. It is securely mounted on the gas cylinder used for assessments or operations to continuously acquire motion and attitude data of the gas cylinder in three-dimensional space.
[0025] Knowledge proficiency score S proThis method objectively measures an operator's mastery of Standard Operating Procedures (SOPs). When candidates verbally recite the SOPs while performing their tasks, the collected acoustic data is first fed into an end-to-end Automatic Speech Recognition (ASR) model based on a Transformer architecture. This ASR model includes a feature extraction layer, a Transformer encoding layer, and a Transformer decoding layer. The feature extraction layer frames and windowes the acoustic data, then transforms it into an acoustic feature vector sequence using a Fast Fourier Transform (FFT). This acoustic feature vector sequence is then input into the Transformer encoding layer, where its multi-head attention mechanism calculates the correlation weight between features at any two time steps in the acoustic feature vector sequence. Even if some pronunciations are masked by the compressor's roar, the attention mechanism can still assign higher weights to key acoustic features based on context (clear pronunciation in the preceding and following text), thus filtering out noise. The Transformer decoding layer receives the high-level acoustic representations extracted by the Transformer encoding layer and performs autoregressive word-by-word prediction, mapping the acoustic features to text to form a text string. This automatic speech recognition model can efficiently and accurately convert speech signals into text strings, and maintain a high recognition rate even in industrial environments with certain background noise. After obtaining the text, text preprocessing is performed, and it is compared with the pre-existing standard SOP text in multiple dimensions. The comparison process includes: (1) Keyword matching, that is, statistically analyzing the coverage of core nouns, operational verbs and key parameters defined in the SOP in the candidate's spoken text. S key (2) Process sequence analysis. In this step, both the SOP and the candidate's oral text are segmented into independent step sentences, and each sentence is converted into a high-dimensional vector using a sentence vector encoding model. Specifically, the sentence vector encoding model can use BERT, which includes: embedding layer, encoding layer, and classification layer. The segmented step sentences are input into the embedding layer and divided into tokens, which are mapped to word vectors, and positional encoding is added. The encoding layer consists of 12 or 24 bidirectional Transformers. Through the self-attention mechanism, each word vector absorbs the context of the entire sentence and outputs a dynamic word vector matrix containing deep semantics. The classification layer includes: fully connected layer, activation layer, normalization layer, and softmax layer. The vector matrix output by the encoding layer flows into the classification layer. After feature extraction, activation, normalization and other processing in each layer of the classification layer, the variable-length word vector matrix is compressed and averaged to integrate into a fixed-length high-dimensional vector. Then, the dynamic time warping (DTW) distance between the two sentence vector sequences is calculated. D DTWTo accurately assess whether the order of the candidate's oral presentation is correct. The DTW algorithm can adapt well to situations such as uneven speaking speed and pauses between steps. (3) Semantic integrity analysis. The overall cosine similarity between the candidate's oral presentation text and the standard SOP text is calculated using a pre-trained language model. Sim cos Specifically, the language model can employ BERT as described above. After pre-training, it extracts a global semantic vector at the full-text level. Then, it compares the global semantic vector A of the candidate's spoken text with the global semantic vector B of the standard SOP text using the cosine similarity formula: Sim cos The formula =(A•B) / (∣A∣×∣B∣) is used to calculate the overall semantic completeness score, allowing for a higher-dimensional assessment of the accuracy of the representation. Finally, the... S key , D DTW , Sim cos Weighted calculation S pro : S pro = α S key + β (1- D DTW )+ γ Sim cos ; α , β , γ These are weighting coefficients adjusted according to the assessment requirements of different positions. The weights are dynamically adjusted based on the job hazard level, ultimately outputting a structured knowledge proficiency score. S pro .
[0026] Psychological quality score S emo Used to gain insight into the operator's inner psychological state through non-content information in speech. From S proA series of temporal acoustic features highly correlated with emotion and physiological tension were extracted from the same acoustic data. These features mainly include: prosodic features reflecting speech rate, such as the mean, variance, and dynamic range of the fundamental frequency (Pitch), as well as speech rate and energy profiles; timbre features reflecting tension, such as acoustic jitter and acoustic flicker, which directly reflect vocal cord muscle tension; and spectral features that macroscopically characterize timbre, such as Mel-frequency cepstral coefficients (MFCCs) and their first and second-order differences. The extracted temporal acoustic features were used to identify emotional states through an emotion recognition model. This emotion recognition model includes: an attention-based gated recurrent unit (GRU), a fully connected layer, and an activation layer. The temporal acoustic features are sequentially processed through the GRU, fully connected layer, and activation layer to output a probability set of emotional state distributions. P_t Specifically, GRU is a special type of recurrent neural network that excels at capturing temporal dependencies in speech signals. Its internal update and reset gate structures enable it to effectively learn long-term dependencies and avoid the vanishing gradient problem. The cell update process of this GRU can be described by the following formula: Update Gate: z _ t = σ ( W _ z [ h _{ t -1}, x _ t ]).
[0027] Reset Door: r _ t = σ ( W _ r [ h _{ t -1}, x _ t ]).
[0028] Candidate hidden state: h '_ t = tanh( W _ h [ r _ t ⊙ h _{ t -1}, x _ t ]).
[0029] Final hidden state: h _t = (1 - z_ t ) ⊙ h _{ t -1} + z _ t ⊙ h '_ t .
[0030] in, h _ t It is the hidden state vector output by the hidden layer at the current time step, which encodes the state from the start of the speech to the current time step. t All time-series characteristic information. x _ t It is the acoustic feature vector input at the current moment. W _ z , W _ r , W _ h These are the weight matrices that the GRU needs to learn at the corresponding positions. σ (·) represents the Sigmoid activation function, and tanh(·) represents the tanh activation function. The tanh activation function introduces non-linearity, enabling the network to learn more complex relationships between inputs and importance. ⊙ represents element-wise multiplication (Hadamard product). During this update process, h '_ t and h _ t The information is processed in a progressive manner to achieve intelligent filtering and memorization of time-series information. h '_ t Its function is to generate candidate hidden states, which represent the state at the current time step. t The gated loop unit is based on the current input x _ t and filtered past information r _ t ⊙ h _{ t -1} calculates a new state that is either a "proposal" or a "candidate". Specifically, it resets the gate. r _ t First, the hidden state of the previous moment was determined. h _{ t How much of the information in `-1}` is relevant to the current calculation? If r _ t If some dimension values are close to 0, it means h _{ t The old information corresponding to `-1` is either "reset" or forgotten because it is irrelevant to generating the current new information. Then, this filtered past information is combined with the current input.x _ t By combining a fully connected layer and a tanh activation function, candidate hidden states are generated. h '_ t .therefore, h '_ t It encapsulates the most important new information at the current moment. h _ t Its function is to calculate and output the final hidden state at the current time step. It is based on the candidate hidden state. h '_ t Based on this, by updating the door z _ t It was ultimately determined through regulation. h _ t The current time step t The output will also serve as the next time step. t +1 input memory. Update gate. z _ t It acts as an "arbitrator," deciding the extent to which the old state from the previous moment should be preserved. h _{ t -1}, and to what extent should the newly generated candidate state be adopted. h '_ t . z _ t The value is between 0 and 1. When its value is close to 1, it means that the GRU decides to adopt more new candidate states. h '_ t When its value is close to 0, it means that the GRU decides to retain more old memories. h _{ t -1}. Through (1 - z_ t ) ⊙ h _{ t -1} and z _ t ⊙ h '_ t This pair of complementary weighted terms, h _ t It achieves a dynamic and smooth integration of past and current information. Therefore, h _ t It is the current moment. tThe final hidden state, rather than an intermediate process, is determined. It precisely summarizes all relevant temporal information up to the current moment and passes it as a memory carrier to the next time step of the GRU, allowing the entire GRU to understand and respond to the complex patterns that change over time in the speech signal. Furthermore, an attention mechanism is introduced into the GRU. This is because not all speech segments in a candidate's complete dictation reflect the candidate's psychological state with equal importance. The attention mechanism can influence the GRU output at different time steps. h _ t Assign different attention weights α_t This allows it to automatically focus on key moments that best reveal emotional changes, such as a sudden tremor in the voice, an unnatural pause, or a sharp change in speech rate. The calculation process for attention weights is as follows: (1) Calculate attention score e _ t : e _ t = tanh( W _ a h _ t + b _ a ).
[0031] Among them, attention score e _ t Hidden state at each time step h _ t The calculated raw score characterizing "importance". e _ t The higher the value, the more crucial emotional information the speech features at that moment may contain. W _ a This represents the weight matrix, which can represent the hidden states of the input. h_t Perform a linear transformation. b _ a The bias vector represents the additional degree of freedom for the linear transformation described above, allowing the GRU to have a basic output even without input, thus increasing the flexibility and fitting ability of the GRU.
[0032] (2) e _ t Normalized to attention weights α _ t : α _ t = softmax( e_t ) = exp( e_t ) / Σi (exp( e_i )).
[0033] Here, exp(·) represents the exponential function, which ensures that all values become positive and amplifies the differences between scores, resulting in higher scores being more likely to be positive. e _ t After exponentialization, it becomes larger, thus dominating the final weights. Only each time point is obtained. e _ t This is insufficient because their scales are inconsistent and cannot be directly used as weights. Therefore, a Softmax function is needed to calculate the scores at all time steps. e _1、 e _2、...、 e _ T It is transformed into a probability distribution with a sum of 1. α _ t The higher the value, the more GRU considers the first... t The more important the voice information at each time step is for judging the overall emotional state, the higher the "attention" it should receive. Next, we will utilize... α _ t To make the final sentiment judgment: During reasoning, GRU first uses the calculated... α _ t For all time steps h_t Perform a weighted summation to generate a single, fixed-length context vector. c : c = Σ t ( α_t h_t) .
[0034] Context vector c It is a highly condensed representation of the entire speech, and it is not a simple average of information from all moments, but rather a key fusion of those moments that the attention mechanism deems most critical. α_t Information about moments with high values. For example, if a test taker's voice suddenly trembles at a certain second, the corresponding... α_t It will be very high, so that one second h_t Features in the final vector c In the middle, it will account for a larger proportion. Then... c The data is fed into a standard fully connected layer (classifier). The classifier then passes through an activation layer and is finally output via the Softmax function. t The probability set of the distribution of emotional states at any given time P_tFor example: {"Confidence": 0.05, "Calmness": 0.10, "Tense": 0.65, "Anxiety": 0.20}. This probability distribution indicates that the most likely emotion expressed in the current audio clip is "tense" (probability 65%). Then, based on the cumulative intensity and duration of negative emotional states (such as tension and anxiety) throughout the assessment process, a comprehensive psychological quality score is calculated. S emo Specifically, the calculation method is as follows: First, the emotional states output by the classifier are classified. These emotional states are divided into two main categories: positive states. C _ pos and negative state C_neg .For example, C _ pos = {Calm, confident} C_neg = {tension, anxiety}. And for each specific c Assign a preset numerical weight w_c Positive state C _ pos The weight is positive, negative state. C_neg The weight is negative, and the absolute value of the weight represents the degree of influence of that state on the final score. For example, it can be set to... w Confidence = 1.0 w Calmness = 0.8 w Tension = -1.2 w Anxiety = -1.5, indicating that the negative impact of anxiety on the score is greater than that of tension. This applies at every time step throughout the entire assessment process. t (For example, every second), a probability set of emotional state distributions will be output. P_t= { P_t ( c_ 1), P_t ( c_ 2), ... , P_t ( c_n Using this P_t and preset w_c Calculate the instantaneous psychological state at that moment. s ( t ): s ( t ) = Σ {c∈C} w_c P_t .
[0035] in, C It is the set of all possible emotional states. s ( tThis can reflect the candidate's psychological state in real time, and its positive or negative value and magnitude directly indicate the degree of quality of the psychological state. Next, in order to comprehensively evaluate the psychological performance throughout the entire assessment process, the instantaneous minutes of all time steps... s ( t By integrating over time, a raw cumulative emotional score can be obtained that simultaneously reflects the "intensity" and "duration" of the psychological state. S_raw Let the total assessment time be... T If there are 1 time steps, then: S_raw = (1 / T ) Σ {t=1 toT} s ( t ).
[0036] Should S_raw The score is a comprehensive indicator; maintaining a positive state for an extended period will cause it to tend towards a positive value, while frequent or prolonged periods of negativity will cause it to tend towards a negative value. Finally, to make the score more readable and standardized, [the following will be implemented / modified]. S_raw A linear mapping is performed to a standardized scoring interval, such as [0, 100], to obtain the final psychological quality score. S emo This mapping process can be defined by the following formula: S emo = S _max ( ( S_raw - S _{ raw \min}) / ( S _{ raw \max} - S _{ raw \min})).
[0037] In the formula, S _max represents the highest score in the standardized scoring range. S _{ raw \max} represents the theoretically best possible raw cumulative sentiment score, which is equal to the maximum weight value in the positive state (e.g., max ( w _confidence, w (Calmness) = 1.0). S _{ raw \min} represents the theoretically worst possible raw cumulative mood score, which is equal to the minimum weight value in the negative state (e.g., min ( w _nervous, wAnxiety () = -1.5). Thus, through the above quantification process, a dynamically changing, multi-dimensional probability sequence of emotional states has been transformed into a single, stable, and physically meaningful comprehensive assessment score. S emo This score not only objectively reflects the average psychological level of the candidates, but also, through its weighted design, imposes a higher penalty on high-risk negative emotions, thus providing unprecedented depth of insight for the selection of personnel for safety-critical positions.
[0038] Operational stability score S sta This method is used to assess candidates' practical skills. First, an Extended Kalman Filter (EKF) data fusion algorithm is used, combined with the gas cylinder's motion attitude data, to calculate the cylinder's precise attitude angles in real time, including pitch, roll, yaw, and acceleration. By monitoring these angles, the total duration for which the gas cylinder's tilt angle exceeds the safe operating range can be recorded. T tilt Next, the smoothness of the operation is analyzed. This is done by calculating the magnitude of the composite acceleration vector. A ( t The system sets an acceleration threshold far exceeding gravitational acceleration. It then detects each impact event during the operation and accumulates the total number of impacts. N ram More precisely, the derivative of acceleration with respect to time, i.e., the impact force, is calculated. J : J ( t ) = d A ( t ) / d t Impact reflects the drastic change in force and is a golden indicator of the smoothness and gentleness of a movement. A smooth operation has a low impact value. This is achieved by calculating the integral of the absolute value of the impact throughout the entire process. | J ( t The smoothness of the overall operation is quantified by |dt. Additionally, the variance of the triaxial angular velocities of the gas cylinder is calculated. Var ( ω To assess the stability of the rotational operation, an operational stability score was calculated. S sta It is a score calculated based on a penalty function algorithm, initially set to full marks, with points deducted based on detected irregularities. S sta = S max -( c 1 N ram +c 2 | J | dt + c 3 T tilt + c 4 ∑ Var ( ω )).
[0039] In the formula, S max This represents the full score. c 1. c 2. c 3. c 4 represents the configurable penalty weight coefficients corresponding to each non-standard operation.
[0040] get S pro , S emo , S sta Then, a weighted fusion is performed to calculate an assessment score that comprehensively reflects the candidate's overall ability. S total : S total = w_p S pro + w_e S emo + w_s S sta .
[0041] In the formula, w_p , w_e , w_s These are the weighting coefficients for the corresponding items, and their sum is 1. The weighting coefficients can be flexibly configured according to the core competency requirements of different positions, and a passing score can be set based on the assessment requirements. If... S total If the score is greater than or equal to the passing score, the operator is considered to have passed the assessment. New employee onboarding and qualification certification can also be conducted. For example, after a batch of new employees completes theoretical training, they need to take a final practical assessment before starting work. If a score of 85 is the passing mark, employees scoring below 85 need to undergo retraining. Furthermore, it is also based on... S pro , S emo , Ssta Provide operators with multi-dimensional competency profiles and specialized training. Examples of application scenarios include: Scenario 1: Employee A and Employee B both scored 88 points and passed the assessment. However, the final score showed: Employee A: S pro It scored 95 points. S sta It was 90 points, but S emo Only 70 points. Analysis shows that when employee A simulated an emergency (such as the system emitting a simulated alarm), his speaking speed increased significantly, his acoustic vibration parameters exceeded the standard, and he made slight impacts when grabbing the gas cylinder.
[0042] Employee B: S pro It is 80 points. S sta It scored 95 points. S emo The score is 95. The analysis shows that employee B's operation was very calm and composed, but he omitted the crucial step of "checking the static grounding" when verbally describing the process.
[0043] Based on the above situation, managers can address the specific issues. Employee A can undergo more stress management and emergency simulation training. Meanwhile, employee B can strengthen their memorization and recitation of Standard Operating Procedures (SOPs). This achieves a shift from "uniform training" to "precise empowerment."
[0044] Scenario 2: Dynamic scheduling and task allocation based on risk grading.
[0045] There is a high-risk task involving the filling of high-pressure, flammable gases. There are also some routine gas filling tasks. The safety supervisor can use the scheduling system to select "top operators" with recent comprehensive evaluation scores above 95 points and no weaknesses in any of the three sub-categories (knowledge, psychology, and operation) to perform the high-risk task. For employees whose "operational stability" scores have fluctuated slightly in recent evaluations, they can be prioritized for lower-risk routine tasks, with reminders given. This achieves risk matching based on data capabilities.
[0046] Scenario 3: Early warning and care for the declining capabilities of veteran employees.
[0047] Operator Wang has 20 years of experience and is highly skilled, but in the two most recent quarterly evaluations, his performance was... S sta (Especially angular velocity variance and impact index) and S emo(Especially in the afternoon) there was a continuous decline. Based on his performance, an automatic trend report was generated and an alert was issued to the safety supervisor. The supervisor wouldn't simply assume Wang's "technical skills have declined," but would proactively intervene and communicate, potentially discovering that Wang hadn't been getting enough rest recently due to family reasons, or was experiencing some health issues. In this way, the company could provide timely humanistic care, adjust his workload, or arrange a medical check-up, eliminating potential safety hazards at their inception and demonstrating the human touch in management.
[0048] Scenario 4: Closed-loop optimization of the training system.
[0049] After a year of data accumulation, it was discovered that over 60% of new employees lost points in the "valve opening sequence" section of the "knowledge proficiency" assessment. Further analysis suggests this may no longer be an issue with individual employees, but rather points to potential deficiencies in the training content or teaching methods. The training department can use this data to focus on optimizing the "valve opening sequence" training module, such as adding animated demonstrations and reinforcing mnemonic devices. The assessment data, in turn, guides and optimizes the training system itself, forming a data-driven, continuous improvement loop.
[0050] In summary, the practical significance of the above comprehensive evaluation results lies in: Objective quantification: transforming operator competence from "gut feeling" to "data-driven assessment." Precise diagnosis: not only knowing the "what" (score levels), but also the "why" (weak areas). Predictive prevention: identifying potential risks and changes in personnel status in advance through trend analysis. Management tool: providing solid data support for management activities such as personnel certification, task allocation, and training optimization.
[0051] II. Real-time risk warning based on acoustic features.
[0052] Acoustic data is collected during the gas cylinder filling process, either during candidate operation assessments or operator daily work. This acoustic data is then input into an acoustic risk identification model to determine the probability distribution of risk categories. Risk categories include "normal," "leakage," and "overpressure." This acoustic risk identification model employs a Convolutional Neural Network (CNN) + Bi-LSTM architecture. First, the acoustic data is converted from a one-dimensional audio waveform signal to a two-dimensional Mel-spectrum. This representation simulates the non-linear auditory characteristics of the human ear, transforming the sound signal into a "fingerprint image" suitable for subsequent image feature extraction. The CNN includes convolutional layers, activation layers, and pooling layers. The convolutional kernels in the CNN scan the Mel-spectrum, automatically learning and extracting local time-frequency features that characterize different acoustic events. For example, the continuous bright band formed in the high-frequency region by the "hissing" sound of a leak, or the brief vertical bright line covering multiple frequencies produced by valve action. Subsequently, the local time-frequency features are input into the activation layer, where they undergo ReLU activation to increase non-linearity, and then max-pooling is performed by the pooling layer to reduce the feature dimensionality, retaining the most significant acoustic fingerprint features. The local time-frequency features, enhanced by a CNN, are then input into a Bidirectional Long Short-Term Memory (Bi-LSTM) network. This Bi-LSTM consists of a forward LTSM, a backward LTSM, a fully connected layer, and an activation layer. The LTSM includes a forget gate to decide whether to retain or discard past information, an input gate to control the update of the current state by the current input, and an output gate to determine the hidden state. Leveraging the unique gating mechanism of LSTM, it can effectively capture the evolution of local time-frequency features over time and long-range contextual relationships. This is crucial for distinguishing between a momentary, harmless metallic knocking sound and a persistent abnormal vibration sound caused by equipment malfunction, even if their spectral characteristics may be similar at a given instant. Finally, the output of the LSTM is fed into the fully connected layer and the activation layer. Through the Softmax activation function, the integrated temporal features are mapped to a probability distribution, explicitly indicating the probability that the current sound belongs to each predefined risk category, such as "normal," "leaking," or "overvoltage."
[0053] The acoustic feature-based real-time risk warning method and the multimodal operator competence assessment method are not used in isolation, but rather form an intelligent closed-loop system. This closed-loop system feeds back the performance of daily risk warnings to the next operator evaluation system, which is the key link in connecting "pre-event assessment" and "in-event monitoring," enabling the entire system to truly integrate from two independent tools into an organic and continuously evolving safety production management ecosystem. The ultimate goal of both assessment and warning is to eliminate risks, achieving "competence assessment -> on-the-job operation -> risk monitoring -> real event feedback -> re-evaluation and optimization." The connection between the two is mainly reflected in two aspects: Relationship 1: The early warning system "provides material" for the evaluation system.
[0054] Data flow: Daily tasks (early warnings) -> Assessment scenarios (evaluation).
[0055] Practical Implications: Imagine that the early warning system detects a rare, slow leak caused by valve aging during routine operation. This real audio data can be added to the assessment system's "case library." In the next assessment, this real leak sound can be played to test whether the operator can identify and react correctly. This transforms the assessment from theoretical to a practical exercise that transcends the real world.
[0056] Related 2: The assessment system "improves the accuracy" of the early warning system.
[0057] Data flow: Assessment scenario (evaluation) -> Daily tasks (early warning).
[0058] Practical significance: In controlled assessment scenarios, multiple data points can be simultaneously collected, including sound, cylinder posture, pressure, and temperature. When a specific impact sound is recorded, the motion posture data can indicate that it's simply a "cylinder gently touching the ground" (safe), rather than a "cylinder falling from a height" (dangerous). This multimodal data with clear scenario annotations can be used to train and optimize a model that relies solely on sound for warnings, informing it that "this sound is not dangerous; do not issue a false alarm." This significantly reduces the false alarm rate of the warning system, making it more reliable in daily operations.
[0059] Furthermore, an interactive interface can be designed based on the gas cylinder filling safety assessment method, including an assessment application for examiners and trainers, and an operation application for frontline operators. The assessment application provides real-time multi-dimensional data visualization, assessment process playback, and the generation and management of detailed evaluation reports. The operation application is deployed on mobile terminals such as smartphones, using a simple and intuitive interface to silently run acoustic risk monitoring in the background, and issuing alarms and operation instructions to operators through sound, vibration, and screen information when necessary. In summary, this invention constructs a complete, data-driven closed-loop safety assurance system by combining an innovative multimodal assessment system with an acoustic risk early warning system based on self-supervised learning. It not only solves the deep-seated defects of traditional assessments and provides a scientific basis for personnel management in high-risk industries, but also transforms advanced artificial intelligence technology into a readily accessible "safety sentinel" for frontline operators, thereby greatly enhancing risk defense capabilities in the pre-event and in-event stages, and has great value in improving the inherent safety level of industrial production.
[0060] It is worth mentioning that the neural networks used in both parts of the reasoning process in the gas cylinder filling safety assessment method require both training and supervised learning. The supervised learning process for the neural networks used in the multimodal operator competence assessment includes: Human Labeling Phase: Before the neural network is officially used, it must undergo a large-scale training process. This process begins with data preparation: first, tens of thousands of audio recordings of test takers are collected. Then, professionals such as psychology experts and experienced interviewers listen to and evaluate each segment of these recordings, assigning a clear emotional label to each segment (for example, a 5-second audio recording is judged as "nervous," while a 10-second recording is judged as "calm"). This label given by the experts is the "standard answer" or "ground truth."
[0061] Model training phase: Massive amounts of expert-labeled data are used to "teach" the model. The model processes a piece of speech and generates its context vector. c The model attempts to make its own predictions. Then, it compares the model's predictions with the "standard answer" given by experts. If a prediction is incorrect, an algorithm called "backpropagation" fine-tunes all the parameters inside the neural network based on the error. This process is repeated millions of times, with each adjustment making the neural network's next prediction closer to the expert's judgment. Ultimately, the classifier learns how to interpret the condensed context vector. c It identifies complex patterns that are highly correlated with specific emotion categories.
[0062] The supervised learning process for neural networks used in real-time risk warning based on acoustic features includes: In a controlled experimental environment, various normal and abnormal filling conditions are accurately simulated, and four heterogeneous data points are simultaneously acquired: high-fidelity audio, high-precision pressure, high-sensitivity valve temperature, and real-time flow. The key to this stage lies in utilizing deterministic physical laws to generate high-confidence labels for nondeterministic acoustic data. For example, based on the Joule-Thomson effect, high-pressure gas leakage will produce a significant cooling effect at the leakage point; therefore, by monitoring the temperature derivative with respect to time, d... T / d t Once its value is less than a preset threshold, the synchronized audio for that time period is automatically marked as a "leakage event". Similarly, when the slope d of the pressure curve... P / d tWhen the slope exceeds the set maximum, the corresponding audio is labeled as an "overpressure event." When pressure and cylinder filling flow data indicate abnormal stagnation or fluctuation, the corresponding audio is labeled as a "blockage event." This method can automatically and efficiently create a large-scale industrial risk acoustic database that covers multiple operating conditions and is highly accurate in its labeling, providing a high-quality data foundation for neural network training. Subsequently, the network parameters are continuously optimized through the backpropagation algorithm to minimize the classification cross-entropy loss between the predicted probability and the true label. L : L = - Σ i Σ d ( y { i , d} log( p { i , d})).
[0063] In the formula, y { i , d} is the true label of one-hot encoding. p { i , d} is a neural network for the first i The sample belongs to the first d The predicted probability of each category. Once the acoustic risk identification model is trained and deployed, its greatest application advantage becomes apparent: it completely liberates monitoring capabilities from physical dependence on specific devices, achieving truly sensorless operation. Once the acoustic risk identification model learns to directly associate specific sound patterns with specific risk events, it no longer requires physical sensors such as pressure and temperature sensors in subsequent practical applications. This characteristic has immense practical value in scenarios with a wide variety of bottle types in assessment projects, allowing for direct operational assessment of any standard, ordinary bottle without the need to purchase, install, debug, and maintain expensive sensors for each assessment object, thus significantly saving hardware, manpower, and time costs. Deployment only requires installing microphones and computing terminals to achieve comprehensive coverage of the assessment area, demonstrating unparalleled flexibility, universality, and economy, ultimately forming a low-cost, highly flexible acoustic risk warning and operational assessment solution.
[0064] After the aforementioned neural network is trained, it is lightweighted and optimized before being deployed to the operator's mobile terminal application. During daily operations, the operator activates the application's "guardian mode," which continuously uses the phone's microphone in the background to collect ambient sound and perform real-time inference locally with extremely low latency. When the probability of a certain risk category output by the neural network continuously exceeds a preset confidence threshold, a tiered warning mechanism is immediately triggered. For example, a low-confidence suspected risk will trigger phone vibration and on-screen text prompts. A medium-confidence risk will be accompanied by clear voice prompts, informing the type of risk and suggesting inspection steps. A high-confidence emergency risk will trigger the highest level of audible and visual alarms and guide the operator to perform emergency shutdown and other operations with clear voice commands.
[0065] In another embodiment, a voice analysis-based risk warning method for gas cylinder filling is also proposed. This method provides acoustic risk warnings during operational assessments using the voice analysis-based gas cylinder filling safety assessment method described in the above embodiments. Specifically, the acoustic risk warning method includes the following steps: Acoustic data was collected during the gas cylinder filling process. This acoustic data was then input into an acoustic risk identification model to obtain the probability distribution of risk categories. The risk categories included: "normal," "leakage," and "overpressure." The identification process of the acoustic risk identification model was as follows: first, the acoustic data was converted into a two-dimensional Mel-frequency spectrogram; then, local time-frequency features representing different acoustic events were extracted from the Mel-frequency spectrogram. These local time-frequency features were processed by an LSTM to capture the evolution of local time-frequency features over time and long-range contextual relationships. Finally, the output of the LSTM was passed through a fully connected layer and an activation function and mapped to the probability distribution of the risk categories. The training of this acoustic risk identification model has been submitted above and will not be elaborated further here.
[0066] In another embodiment, a voice analysis-based gas cylinder filling safety assessment system is proposed, which utilizes the voice analysis-based gas cylinder filling safety assessment method described in the above embodiments. This gas cylinder filling safety assessment system includes: a data acquisition unit, a scoring unit, a judgment unit, a storage unit, and an output unit. The data acquisition unit can employ a microphone array + IMU sensor to collect acoustic data from the operator's verbal description of the operating procedures during the assessment. The scoring unit can employ an embedded chip or a server to receive the acoustic data from the data acquisition unit, compare the acoustic data with the standard operating procedures, and perform accuracy analysis. A knowledge proficiency score is calculated based on the accuracy. S pro Acoustic features are extracted from the acoustic data. These features are then input into an emotion recognition model incorporating an attention mechanism to identify emotional states. Psychological quality scores are calculated based on the duration and cumulative intensity of different emotional states. S emoCollect motion posture data of the gas cylinders. Based on the motion posture data, obtain the duration of excessive tilting. T tilt Number of impacts N ram Impact J angular velocity variance Var ( ω The quantitative indicator group includes [the following]. The operational stability score is calculated based on the quantitative indicator group. S sta .Will S pro , S emo , S sta The weighted fusion is used to obtain the assessment score. S total The judging unit can also be an embedded chip or a server, used to set the passing score based on the assessment requirements. If S total If the score is greater than or equal to the passing score, the operator is deemed to have passed the assessment, and based on... S pro , S emo , S sta Operators are provided with multi-dimensional competency profiles and specialized training. The storage unit can utilize a database to store acoustic data, processing algorithms, and calculation results. The output unit can include a display screen and an alarm, used to display visuals and provide audible and visual alarms based on the output results from the judgment unit.
[0067] In another embodiment, a voice analysis-based safe filling method for gas cylinders is also proposed; please refer to [reference needed]. Figure 2 The system first assesses operators using the voice analysis-based gas cylinder filling safety assessment method described in the above embodiments. Upon passing the assessment, operators are permitted to fill gas cylinders within the assessment period. Simultaneously, while operators are filling gas cylinders, the voice analysis-based gas cylinder filling risk warning method described in the above embodiments is used for acoustic risk warning. When "leakage" or "overpressure" is detected, an alarm is immediately triggered to remind the user to stop the gas cylinder filling operation.
[0068] In another embodiment, a computer-readable storage medium storing computer instructions is also proposed. When the computer instructions are executed by a processor, the steps of the voice analysis-based gas cylinder filling safety assessment method described in the above embodiments are implemented. The computer-readable storage medium may include, but is not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.
[0069] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0070] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.
Claims
1. A method for safety assessment of gas cylinder filling based on voice analysis, characterized in that, It includes: Acoustic data verbally described by operators during operational assessments; Compare acoustic data with standard operating procedures and perform accuracy analysis; The knowledge proficiency score is calculated based on accuracy. S pro ; Acoustic features are extracted from acoustic data; these features are then input into an emotion recognition model incorporating an attention mechanism to identify emotional states; and psychological quality scores are calculated based on the duration and cumulative intensity of different emotional states. S emo ; Collect motion posture data of the gas cylinders; based on the motion posture data, obtain the duration of excessive tilting. T tilt Number of impacts N ram Impact J angular velocity variance Var ( ω The quantitative indicator group includes [the following]; the operational stability score is calculated based on the quantitative indicator group. S sta ; Will S pro , S emo , S sta The weighted fusion is used to obtain the assessment score. S total ; A passing score will be set based on the assessment requirements; like S total If the score is greater than or equal to the passing score, the operator is deemed to have passed the assessment, and based on... S pro , S emo , S sta Operators are given multi-dimensional capability profiles and specialized training.
2. The method for safety assessment of gas cylinder filling based on voice analysis according to claim 1, characterized in that, Perform accuracy analysis and obtain S pro The method includes the following steps: Convert acoustic data into text; The coverage of core nouns, operational verbs, and key parameters in statistical standard operating procedures in the text. S key ; The text and standard operating procedure (SOP) are segmented into independent step sentences, and the DTW distance between the text step sentences and the SOP step sentences is calculated. D DTW ; Calculate the overall cosine similarity between the text and the standard operating procedure. Sim cos ; S key , D DTW , Sim cos Weighted calculation S pro : S pro = α S key + β (1- D DTW )+ γ Sim cos ; α , β , γ These are weighting coefficients that are adjusted according to the performance evaluation requirements of different positions.
3. The method for safety assessment of gas cylinder filling based on voice analysis according to claim 1, characterized in that, acoustic... Features include: prosodic features that reflect speech rate, timbre features that reflect tension, and spectral features that characterize timbre; Perform emotional state recognition and obtain S emo The method includes the following steps: The emotion recognition model includes: an attention-based gated recurrent unit, a fully connected layer, and an activation layer; Acoustic features are first input into a gated recurrent unit (GRU) for feature extraction. While the GRU is processing these features, an attention mechanism is used to output hidden state vectors from different hidden layers within the GRU. h_t Assign different attention weights α_t ; α_t It is used to characterize the importance of the corresponding voice information in judging the overall emotional state; the larger the value, the more important it is. pass α_t Hidden state vectors for all time steps in the gated loop unit h_t Perform weighted summation to generate context vectors c ; Will c The input is obtained from the fully connected layer and the activation layer. t The probability set of the distribution of emotional states at any given time P_t Based on the probability set of emotional state distribution P_t And the preset weight calculation t Instantaneous psychological state at any moment s ( t );right s ( t After time integration and standardization, we obtain S emo .
4. The method for safety assessment of gas cylinder filling based on voice analysis according to claim 3, characterized in that, For different hidden state vectors h_t Assign different attention weights α_t The method includes the following steps: First calculate the time step in each time step h_t Attention score e_t Then e_t After normalization, the corresponding α_t ; in, α_t The calculation formula is: e_t = tanh ( Wa h_t + b_a ); tanh Represents a non-linear activation function. Wa Represents the weight matrix. b_a This represents the bias vector.
5. The method for safety assessment of gas cylinder filling based on voice analysis according to claim 1, characterized in that, The motion attitude data of the gas cylinder includes: pitch angle, roll angle, yaw angle, and acceleration; In the quantitative indicator group, the number of impacts N ram The number of impacts is counted by whether the acceleration of the gas cylinder exceeds the acceleration threshold; if it exceeds the acceleration threshold, the number of impacts is incremented by one. Impact J Calculate the derivative of acceleration with respect to time; S sta The initial score is full marks S max Points will be deducted based on the detected non-standard operations: S sta = S max -( c 1 N ram + c 2 | J | dt + c 3 T tilt + c 4 ∑ Var ( ω )); in, c 1. c 2. c 3. c 4 represents the penalty weighting coefficient for non-standard operations.
6. The method for safety assessment of gas cylinder filling based on voice analysis according to claim 1, characterized in that, Methods for creating multi-dimensional competency profiles and providing specialized training for operators include: When the operator passes the assessment: If S pro If the value is below the threshold, then strengthen the training on memorizing and reciting standard operating procedures; if... S emo If the level is below the threshold, then strengthen psychological stress resistance and emergency simulation training; if S sta If the threshold is not met, then practical training should be strengthened.
7. A method for risk warning during gas cylinder filling based on voice analysis, characterized in that, When using the voice analysis-based gas cylinder filling safety assessment method as described in any one of claims 1 to 6 for operational assessment, it provides acoustic risk warning. The method for acoustic risk warning includes the following steps: Acoustic data is collected during the gas cylinder filling process; the acoustic data is input into an acoustic risk identification model for identification to obtain the probability distribution of risk categories; the risk categories include: "normal", "leakage", and "overpressure"; The acoustic risk identification model's identification process is as follows: First, acoustic data is converted into a two-dimensional Mel spectrogram. Then, local time-frequency features representing different acoustic events are extracted from the Mel spectrogram. The local time-frequency features are processed by LSTM to capture the evolution of local time-frequency features in the time dimension and long-distance contextual relationships. Finally, the output of LSTM is mapped to the probability distribution of risk categories after passing through a fully connected layer and activation function. And / or, all neural network models used in the gas cylinder filling safety assessment methods must be trained before use; the training methods for the acoustic risk identification model include: In a controlled experimental environment, various normal and abnormal filling conditions were simulated, and audio and physical parameters, including pressure, valve temperature and real-time filling flow rate, were collected simultaneously. Based on changes in physical parameters, corresponding risk event labels are assigned to synchronized audio to generate a labeled training dataset; the acoustic risk identification model is then trained using the labeled training dataset. The risk event labels include: when the derivative of pressure with respect to time exceeds the preset maximum slope, the audio corresponding to the same time will be labeled as an overpressure event; when the derivative of valve temperature with respect to time is less than the preset threshold, the audio corresponding to the same time will be labeled as a leakage event; when the pressure and real-time filling flow rate show abnormal stagnation or fluctuation, the audio corresponding to the same time will be labeled as a blockage event.
8. A gas cylinder filling safety assessment system based on voice analysis, characterized in that, It uses the voice analysis-based safety assessment method for gas cylinder filling as described in any one of claims 1 to 6; The gas cylinder filling safety assessment system includes: Acquisition unit, which is used to collect acoustic data verbally described by the operator during the operation assessment; The scoring unit compares acoustic data with standard operating procedures and performs accuracy analysis; based on the accuracy, a knowledge proficiency score is calculated. S pro ; Extract acoustic features from acoustic data; input the acoustic features into an emotion recognition model incorporating an attention mechanism for emotion state recognition; calculate psychological quality scores based on the duration and cumulative intensity of different emotion states. S emo Collect motion posture data of the gas cylinders; based on the motion posture data, obtain the duration of excessive tilting. T tilt Number of impacts N ram Impact J angular velocity variance Var ( ω The quantitative indicator group includes [the following]; the operational stability score is calculated based on the quantitative indicator group. S sta ;Will S pro , S emo , S sta The weighted fusion is used to obtain the assessment score. S total ; The judgment unit is used to set a passing score based on the assessment requirements; if S total If the score is greater than or equal to the passing score, the operator is deemed to have passed the assessment, and based on... S pro , S emo , S sta Operators are given multi-dimensional capability profiles and specialized training.
9. A method for safe filling of gas cylinders based on voice analysis, characterized in that, First, the operator is assessed using the voice analysis-based safety assessment method for gas cylinder filling as described in any one of claims 1 to 6; after passing the assessment, the operator is allowed to fill gas cylinders within the assessment period. When the operator is filling the gas cylinder, the gas cylinder filling risk warning method based on voice analysis as described in claim 7 is used for acoustic risk warning. When "leakage" or "overpressure" is detected, an alarm will be immediately triggered to remind the user to stop the gas cylinder filling operation.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions, which, when executed by a processor, implement the steps of the voice analysis-based gas cylinder filling safety assessment method as described in any one of claims 1 to 6.
Citation Information
Patent Citations
Simulation system for assessing actual gas cylinder filling operations
CN102663915A
Liquid gas cylinder filling operation test simulation system
CN103700310A
Automotive LNG gas cylinder monitoring system and method
CN104329564A
Natural gas pipeline leakage detection method based on LSTM recurrent neural network
CN111022941A
Monitoring system and method for gas storage bottle and fuel cell automobile
CN115060429A
Cited By
Liquefied gas cylinder full-chain traceability and safety early warning method based on multi-modal fusion
CN122222344A
